Skip to content

GitLab

"...composable_kernel_rocm.git" did not exist on "764164b488a9009842c0ce4b14aa74d49eec5e6a"
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
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 654 additions and 278 deletions
example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
...@@ -28,26 +31,26 @@ using ADataType = F16; ...@@ -28,26 +31,26 @@ using ADataType = F16;
using BDataType = F16; using BDataType = F16;
using CDataType = F16; using CDataType = F16;
using ReduceAccDataType = F32; using ReduceAccDataType = F32;
using DDataType = F32; using ReduceDataType = F32;
using DPtrsGlobal = ck::Tuple<DDataType*, DDataType*>; using ReducePtrsGlobal = ck::Tuple<ReduceDataType*, ReduceDataType*>;
using ALayout = ck::tensor_layout::gemm::RowMajor; using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor; using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor; using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough; using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough; using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add; using ReduceOp0 = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add; using ReduceOp1 = ck::reduce::Add;
using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>; using ReduceOps = ck::Tuple<ReduceOp0, ReduceOp1>;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough; using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare; using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>; using ReduceInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>; using ReduceOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
using DGlobalMemOp = using ReduceGlobalMemOps =
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd, ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>; ck::InMemoryDataOperationEnum::AtomicAdd>;
...@@ -60,7 +63,7 @@ using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatc ...@@ -60,7 +63,7 @@ using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatc
//######| | | | 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| //######| | | | 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| //######| | | | | | | | | | | 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>; < 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 // clang-format on
using ReferenceBatchedGemmInstance = ck::tensor_operation::host:: using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
...@@ -140,16 +143,16 @@ int main(int argc, char* argv[]) ...@@ -140,16 +143,16 @@ int main(int argc, char* argv[])
Tensor<CDataType> c_g_m_n_host_result( Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{})); f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>( 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)}))); {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>( 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)}))); {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<CDataType> c_g_m_n_device_result( Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{})); f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>( 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)}))); {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>( 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)}))); {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl; std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
...@@ -174,38 +177,48 @@ int main(int argc, char* argv[]) ...@@ -174,38 +177,48 @@ int main(int argc, char* argv[])
DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace()); DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace()); DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpace()); DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_g_m_device_result.mDesc.GetElementSpace()); DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
DeviceMem d1_device_buf(sizeof(DDataType) * d1_g_m_device_result.mDesc.GetElementSpace()); d0_g_m_device_result.mDesc.GetElementSpace());
DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
d1_g_m_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_g_m_k.mData.data()); a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data()); b_device_buf.ToDevice(b_g_k_n.mData.data());
auto a_element_op = AElementOp{}; auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{}; auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{}; auto c_element_op = CElementOp{};
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()), std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
auto passthrough = UnaryIdenticElementOp{};
auto square = UnarySquareElementOp{};
std::array<void*, 2> reduce_in_element_ops = {&passthrough, &square};
std::array<void*, 2> reduce_out_element_ops = {&passthrough, &passthrough};
std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
reduce1_device_buf.GetDeviceBuffer()};
// do GEMM // do GEMM
auto batched_gemm = DeviceBatchedGemmReduceInstance{}; auto batched_gemm = DeviceBatchedGemmReduceInstance{};
auto invoker = batched_gemm.MakeInvoker(); auto invoker = batched_gemm.MakeInvoker();
auto argument = auto argument = batched_gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
batched_gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()), b_device_buf.GetDeviceBuffer(),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()), nullptr,
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()), {},
dxs_global, c_device_buf.GetDeviceBuffer(),
M, p_reduces,
N, M,
K, N,
StrideA, K,
StrideB, StrideA,
StrideC, StrideB,
a_element_op, StrideC,
b_element_op, {},
c_element_op, gemm_element_ops,
DxsInElementOps{}, {},
DxsOutElementOps{}, reduce_in_element_ops,
BatchCount); reduce_out_element_ops,
BatchCount);
if(!batched_gemm.IsSupportedArgument(argument)) if(!batched_gemm.IsSupportedArgument(argument))
{ {
...@@ -215,8 +228,8 @@ int main(int argc, char* argv[]) ...@@ -215,8 +228,8 @@ int main(int argc, char* argv[])
} }
// init DO, D1 to 0 // init DO, D1 to 0
d0_device_buf.SetZero(); reduce0_device_buf.SetZero();
d1_device_buf.SetZero(); reduce1_device_buf.SetZero();
// if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result // 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 // will not be correct. need to set time_kernel = false for correctness test
...@@ -238,8 +251,8 @@ int main(int argc, char* argv[]) ...@@ -238,8 +251,8 @@ int main(int argc, char* argv[])
if(do_verification) if(do_verification)
{ {
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data()); c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_g_m_device_result.mData.data()); reduce0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_g_m_device_result.mData.data()); reduce1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
auto ref_batched_gemm = ReferenceBatchedGemmInstance{}; auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker(); auto ref_invoker = ref_batched_gemm.MakeInvoker();
...@@ -249,15 +262,15 @@ int main(int argc, char* argv[]) ...@@ -249,15 +262,15 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
auto d0_reduce_op = D0ReduceOp{}; auto reduce0_op = ReduceOp0{};
auto d1_reduce_op = D1ReduceOp{}; auto reduce1_op = ReduceOp1{};
for(int batch = 0; batch < BatchCount; ++batch) for(int batch = 0; batch < BatchCount; ++batch)
{ {
for(int m = 0; m < M; ++m) for(int m = 0; m < M; ++m)
{ {
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>(); auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>(); auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n) for(int n = 0; n < N; ++n)
{ {
...@@ -268,12 +281,12 @@ int main(int argc, char* argv[]) ...@@ -268,12 +281,12 @@ int main(int argc, char* argv[])
UnaryIdenticElementOp{}(d0_val, c_val); UnaryIdenticElementOp{}(d0_val, c_val);
UnarySquareElementOp{}(d1_val, c_val); UnarySquareElementOp{}(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val); reduce0_op(reduce0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val); reduce1_op(reduce1_acc, d1_val);
} }
d0_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d0_acc); d0_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce0_acc);
d1_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d1_acc); d1_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce1_acc);
} }
} }
......
example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <cstdlib> #include <cstdlib>
...@@ -96,15 +99,17 @@ int main() ...@@ -96,15 +99,17 @@ int main()
a_m_n_device_buf.ToDevice(a_m_n.mData.data()); a_m_n_device_buf.ToDevice(a_m_n.mData.data());
b_n_device_buf.ToDevice(b_n.mData.data()); b_n_device_buf.ToDevice(b_n.mData.data());
std::array<const void*, 2> input = {a_m_n_device_buf.GetDeviceBuffer(),
b_n_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_m_n_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides = {Stride, 1};
std::vector<ck::index_t> b_strides = {0, 1};
std::vector<ck::index_t> c_strides = {Stride, 1};
auto broadcastAdd = DeviceElementwiseAddInstance{}; auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(a_m_n_device_buf.GetDeviceBuffer(), auto argument = broadcastAdd.MakeArgumentPointer(
b_n_device_buf.GetDeviceBuffer(), input, output, {M, N}, {a_strides, b_strides}, {c_strides}, Add{});
c_m_n_device_buf.GetDeviceBuffer(),
{M, N},
{Stride, 1},
{0, 1}, // broadcast in first dimension
{Stride, 1},
Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get())) if(!broadcastAdd.IsSupportedArgument(argument.get()))
{ {
......
example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <cstdlib> #include <cstdlib>
...@@ -78,18 +81,24 @@ int main() ...@@ -78,18 +81,24 @@ int main()
a_m_device_buf.ToDevice(a_m.mData.data()); a_m_device_buf.ToDevice(a_m.mData.data());
b_m_n_k_device_buf.ToDevice(b_m_n_k.mData.data()); b_m_n_k_device_buf.ToDevice(b_m_n_k.mData.data());
std::array<const void*, 2> input = {a_m_device_buf.GetDeviceBuffer(),
b_m_n_k_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_m_n_k_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides = {1, 0, 0};
std::vector<ck::index_t> b_strides{b_m_n_k.mDesc.GetStrides().begin(),
b_m_n_k.mDesc.GetStrides().end()};
std::vector<ck::index_t> c_strides{c_m_n_k.mDesc.GetStrides().begin(),
c_m_n_k.mDesc.GetStrides().end()};
auto broadcastAdd = DeviceElementwiseAddInstance{}; auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer( auto argument =
a_m_device_buf.GetDeviceBuffer(), broadcastAdd.MakeArgumentPointer(input,
b_m_n_k_device_buf.GetDeviceBuffer(), output,
c_m_n_k_device_buf.GetDeviceBuffer(), std::vector<ck::index_t>{mnk.begin(), mnk.end()},
std::vector<ck::index_t>{mnk.begin(), mnk.end()}, {a_strides, b_strides},
{1, 0, 0}, // broadcast A on second and third dimension {c_strides},
std::vector<ck::index_t>{b_m_n_k.mDesc.GetStrides().begin(), Add{});
b_m_n_k.mDesc.GetStrides().end()},
std::vector<ck::index_t>{c_m_n_k.mDesc.GetStrides().begin(),
c_m_n_k.mDesc.GetStrides().end()},
Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get())) if(!broadcastAdd.IsSupportedArgument(argument.get()))
{ {
......
example/19_binary_elementwise/elementwise_add_1d.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <cstdlib> #include <cstdlib>
...@@ -76,15 +79,17 @@ int main() ...@@ -76,15 +79,17 @@ int main()
a_m_device_buf.ToDevice(a_m.mData.data()); a_m_device_buf.ToDevice(a_m.mData.data());
b_m_device_buf.ToDevice(b_m.mData.data()); b_m_device_buf.ToDevice(b_m.mData.data());
std::array<const void*, 2> input = {a_m_device_buf.GetDeviceBuffer(),
b_m_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_m_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides = {1};
std::vector<ck::index_t> b_strides = {1};
std::vector<ck::index_t> c_strides = {1};
auto broadcastAdd = DeviceElementwiseAddInstance{}; auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer(a_m_device_buf.GetDeviceBuffer(), auto argument = broadcastAdd.MakeArgumentPointer(
b_m_device_buf.GetDeviceBuffer(), input, output, {M}, {{a_strides}, b_strides}, {c_strides}, Add{});
c_m_device_buf.GetDeviceBuffer(),
{M},
{1},
{1},
{1},
Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get())) if(!broadcastAdd.IsSupportedArgument(argument.get()))
{ {
......
example/19_binary_elementwise/elementwise_add_4d.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <cstdlib> #include <cstdlib>
...@@ -78,16 +81,22 @@ int main() ...@@ -78,16 +81,22 @@ int main()
a_device_buf.ToDevice(a.mData.data()); a_device_buf.ToDevice(a.mData.data());
b_device_buf.ToDevice(b.mData.data()); b_device_buf.ToDevice(b.mData.data());
std::array<const void*, 2> input = {a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {c_device_buf.GetDeviceBuffer()};
std::vector<ck::index_t> a_strides{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()};
std::vector<ck::index_t> b_strides{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()};
std::vector<ck::index_t> c_strides{c.mDesc.GetStrides().begin(), c.mDesc.GetStrides().end()};
auto broadcastAdd = DeviceElementwiseAddInstance{}; auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer( auto argument =
a_device_buf.GetDeviceBuffer(), broadcastAdd.MakeArgumentPointer(input,
b_device_buf.GetDeviceBuffer(), output,
c_device_buf.GetDeviceBuffer(), std::vector<ck::index_t>{nchw.begin(), nchw.end()},
std::vector<ck::index_t>{nchw.begin(), nchw.end()}, {{a_strides}, b_strides},
std::vector<ck::index_t>{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()}, {c_strides},
std::vector<ck::index_t>{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()}, Add{});
std::vector<ck::index_t>{c.mDesc.GetStrides().begin(), c.mDesc.GetStrides().end()},
Add{});
if(!broadcastAdd.IsSupportedArgument(argument.get())) if(!broadcastAdd.IsSupportedArgument(argument.get()))
{ {
......
example/20_convnd_bwd_weight_xdl/convnd_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 <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
......
example/20_convnd_bwd_weight_xdl/convnd_bwd_weight_xdl_bf16_splitk.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
......
example/22_cgemm/cgemm_xdl_fp16.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
......
example/23_softmax/softmax_blockwise.cpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
...@@ -147,6 +150,9 @@ int main(int argc, char* argv[]) ...@@ -147,6 +150,9 @@ int main(int argc, char* argv[])
AccDataType alpha = args.scales[0]; AccDataType alpha = args.scales[0];
AccDataType beta = args.scales[1]; AccDataType beta = args.scales[1];
std::cout << "in: " << in.mDesc << std::endl;
std::cout << "out: " << out.mDesc << std::endl;
std::size_t num_thread = 1; std::size_t num_thread = 1;
if(args.do_verification) if(args.do_verification)
...@@ -192,7 +198,7 @@ int main(int argc, char* argv[]) ...@@ -192,7 +198,7 @@ int main(int argc, char* argv[])
using ReferenceInstance = using ReferenceInstance =
tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>; tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
ReferenceInstance ref; ReferenceInstance ref;
auto ref_arg = ref.MakeArgument(in, out_ref, alpha, beta, Rank, reduceDims); auto ref_arg = ref.MakeArgument(in, out_ref, alpha, beta, reduceDims);
auto invoker = ref.MakeInvoker(); auto invoker = ref.MakeInvoker();
invoker.Run(ref_arg); invoker.Run(ref_arg);
// LogRangeAsType<float>(std::cout << "tensor out_ref: ", out_ref.mData, ",") << std::endl; // LogRangeAsType<float>(std::cout << "tensor out_ref: ", out_ref.mData, ",") << std::endl;
...@@ -209,8 +215,8 @@ int main(int argc, char* argv[]) ...@@ -209,8 +215,8 @@ int main(int argc, char* argv[])
auto argument_ptr = device_instance.MakeArgumentPointer(i_inLengths, auto argument_ptr = device_instance.MakeArgumentPointer(i_inLengths,
i_inStrides, i_inStrides,
reduceDims, reduceDims,
alpha, &alpha,
beta, &beta,
in_dev.GetDeviceBuffer(), in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer()); out_dev.GetDeviceBuffer());
......
example/25_gemm_bias_c_permute/CMakeLists.txt 0 → 100644
View file @ 0b11569f
add_example_executable(example_gemm_bias_c_permute_xdl_fp16 gemm_bias_c_permute_xdl_fp16.cpp)
example/25_gemm_bias_c_permute/gemm_bias_c_permute_xdl_fp16.cpp 0 → 100644
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>
#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_bias_c_permute_xdl.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.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 Add = ck::tensor_operation::element_wise::Add;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DLayout = Row;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmBiasCPermute_Xdl
//######| ALayout| BLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| 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|
//######| | | | Type| Type| Type| DataType| Type| Type| 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| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | 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_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DDataType, EDataType, AElementOp, BElementOp, CDEElementOp, 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<1, 32, 1, 8>, 1>;
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::index_t M0 = 4;
ck::index_t M1 = 32;
ck::index_t M2 = 128;
ck::index_t N0 = 16;
ck::index_t N1 = 256;
// GEMM shape
ck::index_t M = M0 * M1 * M2;
ck::index_t N = N0 * N1;
ck::index_t K = 128;
ck::index_t stride_A = K;
ck::index_t stride_B = K;
#if 1
// E = [M0, N0, M1, N1, M2]
ck::index_t stride_E_M0 = N0 * M1 * N1 * M2;
ck::index_t stride_E_M1 = N1 * M2;
ck::index_t stride_E_M2 = 1;
ck::index_t stride_E_N0 = M1 * N1 * M2;
ck::index_t stride_E_N1 = M2;
// D = [0, N0, 0, N1, 0]
ck::index_t stride_D_M0 = 0;
ck::index_t stride_D_M1 = 0;
ck::index_t stride_D_M2 = 0;
ck::index_t stride_D_N0 = N1;
ck::index_t stride_D_N1 = 1;
#else
// D = [0, 0, 0, N0, N1]
ck::index_t stride_D_M0 = 0;
ck::index_t stride_D_M1 = 0;
ck::index_t stride_D_M2 = 0;
ck::index_t stride_D_N0 = N1;
ck::index_t stride_D_N1 = 1;
// E = [M0, M1, M2, N0, N1]
ck::index_t stride_E_M0 = M1 * M2 * N0 * N1;
ck::index_t stride_E_M1 = M2 * N0 * N1;
ck::index_t stride_E_M2 = N0 * N1;
ck::index_t stride_E_N0 = N1;
ck::index_t stride_E_N1 = 1;
#endif
const ck::tensor_operation::device::DEGridDesc_M0_M1_M2_N0_N1 d_grid_desc{
M0, M1, M2, N0, N1, stride_D_M0, stride_D_M1, stride_D_M2, stride_D_N0, stride_D_N1};
const ck::tensor_operation::device::DEGridDesc_M0_M1_M2_N0_N1 e_grid_desc{
M0, M1, M2, N0, N1, stride_E_M0, stride_E_M1, stride_E_M2, stride_E_N0, stride_E_N1};
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
exit(0);
}
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)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
auto f_host_de_tensor_descriptor =
[](ck::tensor_operation::device::DEGridDesc_M0_M1_M2_N0_N1 de_grid_desc) {
std::size_t m0 = de_grid_desc.M0_;
std::size_t m1 = de_grid_desc.M1_;
std::size_t m2 = de_grid_desc.M2_;
std::size_t n0 = de_grid_desc.N0_;
std::size_t n1 = de_grid_desc.N1_;
std::size_t stride_m0 = de_grid_desc.stride_M0_;
std::size_t stride_m1 = de_grid_desc.stride_M1_;
std::size_t stride_m2 = de_grid_desc.stride_M2_;
std::size_t stride_n0 = de_grid_desc.stride_N0_;
std::size_t stride_n1 = de_grid_desc.stride_N1_;
return HostTensorDescriptor(
std::vector<std::size_t>({m0, m1, m2, n0, n1}),
std::vector<std::size_t>({stride_m0, stride_m1, stride_m2, stride_n0, stride_n1}));
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, stride_A, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, stride_B, BLayout{}));
Tensor<DDataType> d_m0_m1_m2_n0_n1(f_host_de_tensor_descriptor(d_grid_desc));
Tensor<EDataType> e_m0_m1_m2_n0_n1_host_result(f_host_de_tensor_descriptor(e_grid_desc));
Tensor<EDataType> e_m0_m1_m2_n0_n1_device_result(f_host_de_tensor_descriptor(e_grid_desc));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d_m0_m1_m2_n0_n1: " << d_m0_m1_m2_n0_n1.mDesc << std::endl;
std::cout << "e_m0_m1_m2_n0_n1: " << e_m0_m1_m2_n0_n1_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_m0_m1_m2_n0_n1.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_m0_m1_m2_n0_n1.GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
}
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem d_m0_m1_m2_n0_n1_device_buf(sizeof(DDataType) *
d_m0_m1_m2_n0_n1.mDesc.GetElementSpace());
DeviceMem e_m0_m1_m2_n0_n1_device_buf(sizeof(EDataType) *
e_m0_m1_m2_n0_n1_device_result.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
d_m0_m1_m2_n0_n1_device_buf.ToDevice(d_m0_m1_m2_n0_n1.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument = device_op.MakeArgument(a_m_k_device_buf.GetDeviceBuffer(),
b_k_n_device_buf.GetDeviceBuffer(),
d_m0_m1_m2_n0_n1_device_buf.GetDeviceBuffer(),
e_m0_m1_m2_n0_n1_device_buf.GetDeviceBuffer(),
M,
N,
K,
stride_A,
stride_B,
d_grid_desc,
e_grid_desc,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error("wrong! this device_op instance does not support this problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(DDataType) * N + sizeof(EDataType) * M * N;
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, "
<< device_op.GetTypeString() << std::endl;
if(do_verification)
{
Tensor<AccDataType> c_m_n(HostTensorDescriptor(
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m0 = 0; m0 < M0; ++m0)
for(int m1 = 0; m1 < M1; ++m1)
for(int m2 = 0; m2 < M2; ++m2)
for(int n0 = 0; n0 < N0; ++n0)
for(int n1 = 0; n1 < N1; ++n1)
{
int m = m0 * M1 * M2 + m1 * M2 + m2;
int n = n0 * N1 + n1;
cde_element_op(e_m0_m1_m2_n0_n1_host_result(m0, m1, m2, n0, n1),
ck::type_convert<EDataType>(c_m_n(m, n)),
d_m0_m1_m2_n0_n1(m0, m1, m2, n0, n1));
}
e_m0_m1_m2_n0_n1_device_buf.FromDevice(e_m0_m1_m2_n0_n1_device_result.mData.data());
return ck::utils::check_err(e_m0_m1_m2_n0_n1_device_result.mData,
e_m0_m1_m2_n0_n1_host_result.mData)
? 0
: 1;
}
return 0;
}
example/CMakeLists.txt
View file @ 0b11569f
...@@ -43,3 +43,4 @@ add_subdirectory(21_gemm_layernorm) ...@@ -43,3 +43,4 @@ add_subdirectory(21_gemm_layernorm)
add_subdirectory(22_cgemm) add_subdirectory(22_cgemm)
add_subdirectory(23_softmax) add_subdirectory(23_softmax)
add_subdirectory(24_batched_gemm_c_permute) add_subdirectory(24_batched_gemm_c_permute)
add_subdirectory(25_gemm_bias_c_permute)
include/ck/device_utility/device_prop.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
#include <string> #include <string>
......
include/ck/device_utility/hip_check_error.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
#include <hip/hip_runtime.h> #include <hip/hip_runtime.h>
......
include/ck/device_utility/kernel_launch.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once #pragma once
#include <hip/hip_runtime.h> #include <hip/hip_runtime.h>
......
include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1_nhwc_kyxc_nhwk.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1_NHWC_KYXC_NHWK_HPP #ifndef CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1_NHWC_KYXC_NHWK_HPP
#define CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1_NHWC_KYXC_NHWK_HPP #define CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1_NHWC_KYXC_NHWK_HPP
......
include/ck/problem_transform/transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1R2_NHWC_KYXC_NHWK_HPP #ifndef CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1R2_NHWC_KYXC_NHWK_HPP
#define CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1R2_NHWC_KYXC_NHWK_HPP #define CK_TRANSFORM_BACKWARD_DATA_CONVOLUTION_INTO_GEMM_V4R1R2_NHWC_KYXC_NHWK_HPP
......
include/ck/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP #ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP
#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP #define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP
......
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