Skip to content

GitLab

Commit 0b11569f authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into batched_gemm_c_permute

parents e8d3a0fb fa9a0a5c
Hide whitespace changes
Inline Side-by-side
Showing with 181 additions and 109 deletions
example/03_gemm_bias_relu/gemm_xdl_bias_relu.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
example/06_conv2d_fwd_bias_relu/conv2d_fwd_xdl_bias_relu.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
example/07_conv2d_fwd_bias_relu_add/conv2d_fwd_xdl_bias_relu_add.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <numeric>
......
example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <numeric>
......
example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <numeric>
......
example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <numeric>
......
example/10_conv2d_bwd_data/conv2d_bwd_data_xdl.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
example/11_conv2d_bwd_weight/conv2d_bwd_weight_xdl.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
example/12_reduce/reduce_blockwise.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
example/12_reduce/reduce_blockwise_two_call.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <sstream>
......
example/13_pool2d_fwd/pool2d_fwd_common.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
......
example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <cstdlib>
......
example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <cstdlib>
......
example/14_gemm_xdl_requant_relu_requant/gemm_xdl_requant_relu_requant_int8.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......@@ -30,19 +33,19 @@ using BDataType = F16;
using CDataType = F16;
using GemmAccDataType = F32;
using ReduceAccDataType = F32;
using DDataType = F64;
using DPtrsGlobal = ck::Tuple<DDataType*>;
using ReduceDataType = F64;
using ReducePtrsGlobal = ck::Tuple<ReduceDataType*>;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using DsReduceOp = ck::Tuple<ck::reduce::Max>;
using DsElementOp = ck::Tuple<ck::tensor_operation::element_wise::PassThrough>;
using DGlobalMemOp =
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using ReduceOps = ck::Tuple<ck::reduce::Max>;
using ReduceElementOps = ck::Tuple<ck::tensor_operation::element_wise::PassThrough>;
using ReduceGlobalMemOps =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
static constexpr auto GemmSpecialization =
......@@ -50,11 +53,11 @@ static constexpr auto GemmSpecialization =
// clang-format off
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| 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| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| 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| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, ReduceAccDataType, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DsReduceOp, DsElementOp, DsElementOp, DGlobalMemOp, GemmSpecialization, 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<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| ReduceData| A| B| C| Reduce| ReduceInEleOp| ReduceAccEleOp| Reduce| 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| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Operation| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | 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| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, ReduceAccDataType, ReducePtrsGlobal, AElementOp, BElementOp, CElementOp, ReduceOps, ReduceElementOps, ReduceElementOps, ReduceGlobalMemOps, GemmSpecialization, 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<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
......@@ -65,12 +68,12 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp
BElementOp,
CElementOp>;
template <typename ADataType, typename BDataType, typename CDataType, typename DDataType>
template <typename ADataType, typename BDataType, typename CDataType, typename ReduceDataType>
void DumpGemmLayerNormPerf(float gemm_reduce_time, int M, int N, int K)
{
std::size_t gemm_flop = std::size_t(2) * M * N * K;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(DDataType) * M;
sizeof(CDataType) * M * N + sizeof(ReduceDataType) * M;
float tflops = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / gemm_reduce_time;
......@@ -145,17 +148,17 @@ int main(int argc, char* argv[])
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d_m_host_result(
Tensor<ReduceDataType> reduce_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d_m_device_result(
Tensor<ReduceDataType> reduce_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "d_m: " << d_m_host_result.mDesc << std::endl;
std::cout << "reduce_m: " << reduce_m_host_result.mDesc << std::endl;
switch(init_method)
{
......@@ -173,35 +176,40 @@ int main(int argc, char* argv[])
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem d_device_buf(sizeof(DDataType) * d_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce_device_buf(sizeof(ReduceDataType) *
reduce_m_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto ds_element_op = DsElementOp{};
auto p_ds_global = ck::make_tuple(static_cast<DDataType*>(d_device_buf.GetDeviceBuffer()));
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto reduce_element_op = ReduceElementOps{}[ck::Number<0>{}];
std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
std::array<void*, 1> reduce_element_ops = {&reduce_element_op};
std::array<void*, 1> p_reduces = {reduce_device_buf.GetDeviceBuffer()};
// do GEMM
auto gemm = DeviceGemmReduceInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
p_ds_global,
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
nullptr,
{},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
ds_element_op,
ds_element_op);
{},
gemm_element_ops,
{},
reduce_element_ops,
reduce_element_ops);
if(!gemm.IsSupportedArgument(argument))
{
......@@ -212,7 +220,7 @@ int main(int argc, char* argv[])
// [CAUSION]: launch_and_time_kernel will not initialize D.
// If we evaluate kernel multiple time but without initialize D. Verification will fail
d_device_buf.SetValue(ck::NumericLimits<DDataType>::Lowest());
reduce_device_buf.SetValue(ck::NumericLimits<ReduceDataType>::Lowest());
invoker.Run(argument, StreamConfig{nullptr, false});
bool pass = true;
......@@ -220,7 +228,7 @@ int main(int argc, char* argv[])
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
d_device_buf.FromDevice(d_m_device_result.mData.data());
reduce_device_buf.FromDevice(reduce_m_device_result.mData.data());
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -230,27 +238,27 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
auto d_reduce_op = DsReduceOp{}[ck::Number<0>{}];
auto reduce_op = ReduceOps{}[ck::Number<0>{}];
for(int m = 0; m < M; ++m)
{
ReduceAccDataType d_acc = d_reduce_op.GetIdentityValue<ReduceAccDataType>();
ReduceAccDataType reduce_acc = reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType curr_val =
ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
d_reduce_op(d_acc, curr_val);
reduce_op(reduce_acc, curr_val);
};
d_m_host_result(m) = d_acc;
reduce_m_host_result(m) = reduce_acc;
}
pass = ck::utils::check_err(c_m_n_device_result.mData,
c_m_n_host_result.mData,
"Error: Incorrect results c") &&
ck::utils::check_err(d_m_device_result.mData,
d_m_host_result.mData,
ck::utils::check_err(reduce_m_device_result.mData,
reduce_m_host_result.mData,
"Error: Incorrect results d",
1e-3,
1e-3);
......@@ -260,7 +268,7 @@ int main(int argc, char* argv[])
{
float gemm_reduceMax_ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, DDataType>(
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, ReduceDataType>(
gemm_reduceMax_ave_time, M, N, K);
}
......
example/16_gemm_reduce/gemm_reduce_xdl_mean_squaremean_fp16.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......@@ -30,27 +33,27 @@ using BDataType = F16;
using CDataType = F16;
using GemmAccDataType = F32;
using ReduceAccDataType = F32;
using DDataType = F32;
using DPtrsGlobal = ck::Tuple<DDataType*, DDataType*>;
using ReduceDataType = F32;
using ReducePtrsGlobal = ck::Tuple<ReduceDataType*, ReduceDataType*>;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using ReduceOp0 = ck::reduce::Add;
using ReduceOp1 = ck::reduce::Add;
using ReduceOps = ck::Tuple<ReduceOp0, ReduceOp1>;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using ReduceInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using ReduceOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using DGlobalMemOp =
using ReduceGlobalMemOps =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
......@@ -59,11 +62,11 @@ static constexpr auto GemmSpecialization =
// clang-format off
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| 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| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | Operation| 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| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOps, DxsOutElementOps, DGlobalMemOp, GemmSpecialization, 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<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| ReduceDData| A| B| C| Reduce| ReduceInEleOp| ReduceOutEleOp| Reduce| 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| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Operation| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//######| | | | | | | | | | | 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| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, ReducePtrsGlobal, AElementOp, BElementOp, CElementOp, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceGlobalMemOps, GemmSpecialization, 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<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
......@@ -74,13 +77,13 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp
BElementOp,
CElementOp>;
template <typename ADataType, typename BDataType, typename CDataType, typename DDataType>
template <typename ADataType, typename BDataType, typename CDataType, typename ReduceDataType>
void DumpGemmLayerNormPerf(float gemm_reduce_time, int M, int N, int K)
{
std::size_t gemm_flop = std::size_t(2) * M * N * K;
std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(DDataType) * M +
sizeof(DDataType) * M;
sizeof(CDataType) * M * N + sizeof(ReduceDataType) * M +
sizeof(ReduceDataType) * M;
float tflops = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
float gemm_gb_per_sec = gemm_num_byte / 1.E6 / gemm_reduce_time;
......@@ -155,22 +158,22 @@ int main(int argc, char* argv[])
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_host_result(
Tensor<ReduceDataType> reduce0_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_host_result(
Tensor<ReduceDataType> reduce1_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_device_result(
Tensor<ReduceDataType> reduce0_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_device_result(
Tensor<ReduceDataType> reduce1_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
switch(init_method)
{
......@@ -188,39 +191,48 @@ int main(int argc, char* argv[])
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
reduce0_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
reduce1_m_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
auto passthrough = UnaryIdenticElementOp{};
auto square = UnarySquareElementOp{};
auto div = UnaryDivElementOp{N};
std::array<void*, 2> reduce_in_element_ops = {&passthrough, &square};
std::array<void*, 2> reduce_out_element_ops = {&div, &div};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{N, N};
std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
reduce1_device_buf.GetDeviceBuffer()};
// do GEMM
auto gemm = DeviceGemmReduceInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
dxs_global,
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
nullptr,
{},
c_device_buf.GetDeviceBuffer(),
p_reduces,
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
dxs_in_element_op,
dxs_out_element_op);
{},
gemm_element_ops,
{},
reduce_in_element_ops,
reduce_out_element_ops);
if(!gemm.IsSupportedArgument(argument))
{
......@@ -229,9 +241,9 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
// init reducetion buffer to 0
reduce0_device_buf.SetZero();
reduce1_device_buf.SetZero();
// if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
// will not be correct. need to set time_kernel = false for correctness test
......@@ -241,8 +253,8 @@ int main(int argc, char* argv[])
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_m_device_result.mData.data());
reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -252,42 +264,40 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
auto d0_reduce_op = D0ReduceOp{};
auto d1_reduce_op = D1ReduceOp{};
auto reduce0_op = ReduceOp0{};
auto reduce1_op = ReduceOp1{};
for(int m = 0; m < M; ++m)
{
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
auto c_val = ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
ReduceAccDataType square_c_val;
square(square_c_val, c_val);
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
reduce0_op(reduce0_acc, c_val);
reduce1_op(reduce1_acc, square_c_val);
}
dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
div(reduce0_acc, reduce0_acc);
div(reduce1_acc, reduce1_acc);
reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
}
pass = ck::utils::check_err(c_m_n_device_result.mData,
c_m_n_host_result.mData,
"Error: Incorrect results c") &&
ck::utils::check_err(d0_m_device_result.mData,
d0_m_host_result.mData,
ck::utils::check_err(reduce0_m_device_result.mData,
reduce0_m_host_result.mData,
"Error: Incorrect results d0",
1e-4,
1e-5) &&
ck::utils::check_err(d1_m_device_result.mData,
d1_m_host_result.mData,
ck::utils::check_err(reduce1_m_device_result.mData,
reduce1_m_host_result.mData,
"Error: Incorrect results d1",
1e-3,
1e-5);
......@@ -297,7 +307,7 @@ int main(int argc, char* argv[])
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, DDataType>(ave_time, M, N, K);
DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, ReduceDataType>(ave_time, M, N, K);
}
return pass ? 0 : 1;
......
example/17_convnd_bwd_data_xdl/convnd_bwd_data_xdl.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
......
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