ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 10:09:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fused GEMM+GEMM (#351)

Browse Source 
* initial stub for gemm_gemm_xdl_cshuffle

* set up example code

* compiles

* prevent integer overflow

* harmonize interface between ref_gemm and ref_batched_gemm

* batched_gemm_gemm

* fix example

* host tensor gen: diagonal pattern in lowest two-dimensions only

* make c descriptors containing only integral constants

* clean up

* add BlockwiseGemmXdlops_v2 while exploring an unified approach

* implement proper interface

* tidy up example

* fix compilation warnings

* coarsely controlled 2nd gemm padding

* remove rocm-cmake's hard requirement for certain revision

* clang-format

* resolve merge conflict

* fix compilation error on gfx10

* adds acc0 elementwise op to interface

* add gemm_gemm instances and tests

* avoid LDS data hazard

* fix build

Co-authored-by: Chao Liu <chao.liu2@amd.com>

[ROCm/composable_kernel commit: c20a75b07d]
This commit is contained in:
Anthony Chang
2022-08-13 22:18:58 +08:00
committed by GitHub
parent 4dab000987
commit bd30eaf33b
32 changed files with 2822 additions and 689 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
View File

@@ -10,9 +10,9 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.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/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
template <ck::index_t... Is>
@@ -186,9 +186,9 @@ int main(int argc, char* argv[])
b_k_n.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
}
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 c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data());

2
example/30_grouped_convnd_fwd_bias_relu/CMakeLists.txt
View File

@@ -1,2 +0,0 @@
add_example_executable(example_grouped_convnd_fwd_bias_relu_xdl_fp16 grouped_convnd_fwd_bias_relu_xdl_fp16.cpp)
target_link_libraries(example_grouped_convnd_fwd_bias_relu_xdl_fp16 PRIVATE utility)

28
example/30_grouped_convnd_fwd_bias_relu/README.md
View File

@@ -1,28 +0,0 @@
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: time kernel (0=no, 1=yes)
#Following arguments (depending on number of spatial dims):
# N spatial dimensions
# G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D)
# <strides>, (ie Sy, Sx for 2D)
# <dilations>, (ie Dy, Dx for 2D)
# <left padding>, (ie LeftPy, LeftPx for 2D)
# <right padding>, (ie RightPy, RightPx for 2D)
bin/example_grouped_convnd_fwd_bias_relu_xdl_fp16 1 1 1
```
Result (MI100)
```
in: dim 5, lengths {1, 128, 192, 71, 71}, strides {6912, 967872, 1, 13632, 192}
wei: dim 5, lengths {1, 256, 192, 3, 3}, strides {192, 1728, 1, 576, 192}
bias: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
out: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 331776, 1, 9216, 256}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 1.19215 ms, 123.112 TFlops, 279.827 GB/s, DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<256, 128, 256, 32, Default>
```

192
example/30_grouped_convnd_fwd_bias_relu/grouped_convnd_fwd_bias_common.hpp
View File

@@ -1,192 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <type_traits>
#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/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
void print_helper_msg()
{
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;
}
template <ck::index_t NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp,
typename DeviceConvNDFwdInstance>
int run_grouped_conv_fwd_bias(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& bias_g_n_k_wos_desc,
const HostTensorDescriptor& out_g_n_k_wos_desc,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
{
Tensor<InDataType> in(in_g_n_c_wis_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
Tensor<OutDataType> bias(bias_g_n_k_wos_desc);
Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
std::cout << "in: " << in.mDesc << std::endl;
std::cout << "wei: " << wei.mDesc << std::endl;
std::cout << "bias: " << bias.mDesc << std::endl;
std::cout << "out: " << out_host.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
bias.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break;
default:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
bias.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
}
DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
DeviceMem bias_device_buf(sizeof(OutDataType) * bias.mDesc.GetElementSpaceSize());
DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
in_device_buf.ToDevice(in.mData.data());
wei_device_buf.ToDevice(wei.mData.data());
bias_device_buf.ToDevice(bias.mData.data());
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
copy(bias_g_n_k_wos_desc.GetLengths(), d_g_n_k_wos_lengths);
copy(bias_g_n_k_wos_desc.GetStrides(), d_g_n_k_wos_strides);
copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads);
copy(conv_param.input_right_pads_, input_right_pads);
// do Conv
auto conv = DeviceConvNDFwdInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(
in_device_buf.GetDeviceBuffer(),
wei_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
out_device_buf.GetDeviceBuffer(),
a_g_n_c_wis_lengths,
a_g_n_c_wis_strides,
b_g_k_c_xs_lengths,
b_g_k_c_xs_strides,
std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_lengths}},
std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_strides}},
e_g_n_k_wos_lengths,
e_g_n_k_wos_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
if(!conv.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem");
}
float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = conv_param.GetFlops();
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
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;
if(do_verification)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
Tensor<OutDataType> c_host(out_g_n_k_wos_desc);
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
PassThrough>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in,
wei,
c_host,
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,
PassThrough{});
ref_invoker.Run(ref_argument);
// TODO: implement elementwise operation for host
out_host.ForEach(
[&](auto&, auto idx) { out_element_op(out_host(idx), c_host(idx), bias(idx)); });
out_device_buf.FromDevice(out_device.mData.data());
return ck::utils::check_err(
out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f)
? 0
: 1;
}
return 0;
}

437
example/30_grouped_convnd_fwd_bias_relu/grouped_convnd_fwd_bias_relu_xdl_fp16.cpp
View File

@@ -1,437 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "grouped_convnd_fwd_bias_common.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using BiasDataType = ck::half_t;
using OutDataType = ck::half_t;
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::AddRelu;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
#if 1
template <ck::index_t NDimSpatial,
typename InLayout,
typename WeiLayout,
typename BiasLayout,
typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<BiasLayout>,
OutLayout,
InDataType,
WeiDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<BiasDataType>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
32, // KPerBlock
8, // AK1
8, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1,
1,
S<1, 32, 1, 8>,
8>;
#else
template <ck::index_t NDimSpatial,
typename InLayout,
typename WeiLayout,
typename BiasLayout,
typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<BiasLayout>,
OutLayout,
InDataType,
WeiDataType,
AccDataType,
CShuffleDataType,
ck::Tuple<BiasDataType>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
256, // MPerBlock
16, // NPerBlock
32, // KPerBlock
8, // AK1
8, // BK1
16, // MPerXdl
16, // NPerXdl
4, // MXdlPerWave
1, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 16, 4>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
2, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
4, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 256, 1, 1>,
1>;
#endif
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;
// conventional group conv definition
// G = 2
// [N, C, Hi, Wi] = [128, 384, 71, 71]
// [K, C, Y, X] = [512, 192, 3, 3]
// [N, K, Ho, Wo] = [128, 512, 36, 36]
// CK group conv definition
// [G, N, C, Hi, Wi] = [2, 128, 192, 71, 71]
// [G, K, C, Y, X] = [2, 256, 192, 3, 3]
// [G, N, K, Ho, Wo] = [2, 128, 256, 36, 36]
ck::utils::conv::ConvParam conv_param{
2, 2, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
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);
}
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::G_NW_C;
using WeiLayout = ctc::G_K_X_C;
using BiasLayout = ctc::G_NW_K;
using OutLayout = ctc::G_NW_K;
const auto in_g_n_c_wis_desc = HostTensorDescriptor(
{conv_param.G_, conv_param.N_, conv_param.C_, conv_param.input_spatial_lengths_[0]},
{
conv_param.C_, // g
conv_param.input_spatial_lengths_[0] * conv_param.G_ * conv_param.C_, // n
1, // c
conv_param.G_ * conv_param.C_ // wi
});
const auto wei_g_k_c_xs_desc = HostTensorDescriptor(
{conv_param.G_, conv_param.K_, conv_param.C_, conv_param.filter_spatial_lengths_[0]},
{
conv_param.K_ * conv_param.filter_spatial_lengths_[0] * conv_param.C_, // g
conv_param.filter_spatial_lengths_[0] * conv_param.C_, // k
1, // c
conv_param.C_ // x
});
const auto bias_g_n_k_wos_desc = HostTensorDescriptor(
{conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
{
conv_param.K_, // g
0, // k
1, // c
0 // x
});
const auto out_g_n_k_wos_desc = HostTensorDescriptor(
{conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
{
conv_param.K_, // g
conv_param.output_spatial_lengths_[0] * conv_param.G_ * conv_param.K_, // n
1, // k
conv_param.G_ * conv_param.K_ // wo
});
return run_grouped_conv_fwd_bias<
1,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<1, InLayout, WeiLayout, BiasLayout, OutLayout>>(
do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
bias_g_n_k_wos_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 2)
{
using InLayout = ctc::G_NHW_C;
using WeiLayout = ctc::G_K_YX_C;
using BiasLayout = ctc::G_NHW_K;
using OutLayout = ctc::G_NHW_K;
const auto in_g_n_c_wis_desc = HostTensorDescriptor(
{conv_param.G_,
conv_param.N_,
conv_param.C_,
conv_param.input_spatial_lengths_[0],
conv_param.input_spatial_lengths_[1]},
{
conv_param.C_, // g
conv_param.input_spatial_lengths_[0] * conv_param.input_spatial_lengths_[1] *
conv_param.G_ * conv_param.C_, // n
1, // c
conv_param.input_spatial_lengths_[1] * conv_param.G_ * conv_param.C_, // hi
conv_param.G_ * conv_param.C_ // wi
});
const auto wei_g_k_c_xs_desc =
HostTensorDescriptor({conv_param.G_,
conv_param.K_,
conv_param.C_,
conv_param.filter_spatial_lengths_[0],
conv_param.filter_spatial_lengths_[1]},
{
conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
conv_param.filter_spatial_lengths_[1] * conv_param.C_, // g
conv_param.filter_spatial_lengths_[0] *
conv_param.filter_spatial_lengths_[1] * conv_param.C_, // k
1, // c
conv_param.filter_spatial_lengths_[1] * conv_param.C_, // y
conv_param.C_ // x
});
const auto bias_g_n_k_wos_desc =
HostTensorDescriptor({conv_param.G_,
conv_param.N_,
conv_param.K_,
conv_param.output_spatial_lengths_[0],
conv_param.output_spatial_lengths_[1]},
{
conv_param.K_, // g
0, // n
1, // k
0, // ho
0 // wo
});
const auto out_g_n_k_wos_desc = HostTensorDescriptor(
{conv_param.G_,
conv_param.N_,
conv_param.K_,
conv_param.output_spatial_lengths_[0],
conv_param.output_spatial_lengths_[1]},
{
conv_param.K_, // g
conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
conv_param.G_ * conv_param.K_, // n
1, // k
conv_param.output_spatial_lengths_[1] * conv_param.G_ * conv_param.K_, // ho
conv_param.G_ * conv_param.K_ // wo
});
return run_grouped_conv_fwd_bias<
2,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<2, InLayout, WeiLayout, BiasLayout, OutLayout>>(
do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
bias_g_n_k_wos_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 3)
{
using InLayout = ctc::G_NDHW_C;
using WeiLayout = ctc::G_K_ZYX_C;
using BiasLayout = ctc::G_NDHW_K;
using OutLayout = ctc::G_NDHW_K;
const auto in_g_n_c_wis_desc = HostTensorDescriptor(
{conv_param.G_,
conv_param.N_,
conv_param.C_,
conv_param.input_spatial_lengths_[0],
conv_param.input_spatial_lengths_[1],
conv_param.input_spatial_lengths_[2]},
{
conv_param.C_, // g
conv_param.input_spatial_lengths_[0] * conv_param.input_spatial_lengths_[1] *
conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // n
1, // c
conv_param.input_spatial_lengths_[1] * conv_param.input_spatial_lengths_[2] *
conv_param.G_ * conv_param.C_, // di
conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // hi
conv_param.G_ * conv_param.C_ // wi
});
const auto wei_g_k_c_xs_desc = HostTensorDescriptor(
{conv_param.G_,
conv_param.K_,
conv_param.C_,
conv_param.filter_spatial_lengths_[0],
conv_param.filter_spatial_lengths_[1],
conv_param.filter_spatial_lengths_[2]},
{
conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
conv_param.filter_spatial_lengths_[1] * conv_param.filter_spatial_lengths_[2] *
conv_param.C_, // g
conv_param.filter_spatial_lengths_[0] * conv_param.filter_spatial_lengths_[1] *
conv_param.filter_spatial_lengths_[2] * conv_param.C_, // k
1, // c
conv_param.filter_spatial_lengths_[1] * conv_param.filter_spatial_lengths_[2] *
conv_param.C_, // z
conv_param.filter_spatial_lengths_[2] * conv_param.C_, // y
conv_param.C_ // x
});
const auto bias_g_n_k_wos_desc =
HostTensorDescriptor({conv_param.G_,
conv_param.N_,
conv_param.K_,
conv_param.output_spatial_lengths_[0],
conv_param.output_spatial_lengths_[1],
conv_param.output_spatial_lengths_[2]},
{
conv_param.K_, // g
0, // n
1, // k
0, // z
0, // y
0 // x
});
const auto out_g_n_k_wos_desc = HostTensorDescriptor(
{conv_param.G_,
conv_param.N_,
conv_param.K_,
conv_param.output_spatial_lengths_[0],
conv_param.output_spatial_lengths_[1],
conv_param.output_spatial_lengths_[2]},
{
conv_param.K_, // g
conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // n
1, // k
conv_param.output_spatial_lengths_[1] * conv_param.output_spatial_lengths_[2] *
conv_param.G_ * conv_param.K_, // do
conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // ho
conv_param.G_ * conv_param.K_ // wo
});
return run_grouped_conv_fwd_bias<
3,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<3, InLayout, WeiLayout, BiasLayout, OutLayout>>(
do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
bias_g_n_k_wos_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
return 0;
}

0
example/31_grouped_convnd_fwd_bias_relu_add/CMakeLists.txt → example/30_grouped_convnd_fwd_bias_relu_add/CMakeLists.txt
View File

0
example/31_grouped_convnd_fwd_bias_relu_add/README.md → example/30_grouped_convnd_fwd_bias_relu_add/README.md
View File

0
example/31_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_common.hpp → example/30_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_common.hpp
View File

0
example/31_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_xdl_bf16.cpp → example/30_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_xdl_bf16.cpp
View File

0
example/31_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_xdl_fp16.cpp → example/30_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_xdl_fp16.cpp
View File

0
example/31_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_xdl_fp32.cpp → example/30_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_xdl_fp32.cpp
View File

0
example/31_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_xdl_int8.cpp → example/30_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_xdl_int8.cpp
View File

1
example/31_batched_gemm_gemm/CMakeLists.txt Normal file
View File

@@ -0,0 +1 @@
add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)

371
example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp Normal file
View File

@@ -0,0 +1,371 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
|------------|
Gemm0
|---------------------|
Gemm1
*/
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm_xdl_cshuffle.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/reference_tensor_operation/cpu/reference_batched_gemm.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 ADataType = F16;
using B0DataType = F16;
using B1DataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = F16;
using ALayout = Row;
using B0Layout = Col;
using B1Layout = Row;
using CLayout = Row;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = PassThrough;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
ALayout,
B0Layout,
B1Layout,
CLayout,
ADataType,
B0DataType,
B1DataType,
CDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmDefault,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
128, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
4, // Gemm1NXdlPerWave
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<8, 32, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CShuffleBlockTransferScalarPerVector_NPerBlock
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
ADataType,
AccDataType,
AElementOp,
B0ElementOp,
CElementOp>;
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
AccDataType,
AElementOp,
B1ElementOp,
CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t K = 64;
ck::index_t O = 128;
ck::index_t BatchCount = 4;
ck::index_t StrideA = -1;
ck::index_t StrideB0 = -1;
ck::index_t StrideB1 = -1;
ck::index_t StrideC = -1;
ck::index_t BatchStrideA = -1;
ck::index_t BatchStrideB0 = -1;
ck::index_t BatchStrideB1 = -1;
ck::index_t BatchStrideC = -1;
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 if(argc == 9)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
BatchCount = std::stoi(argv[8]);
}
else if(argc == 17)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
BatchCount = std::stoi(argv[8]);
StrideA = std::stoi(argv[9]);
StrideB0 = std::stoi(argv[10]);
StrideB1 = std::stoi(argv[11]);
StrideC = std::stoi(argv[12]);
BatchStrideA = std::stoi(argv[13]);
BatchStrideB0 = std::stoi(argv[14]);
BatchStrideB1 = std::stoi(argv[15]);
BatchStrideC = std::stoi(argv[16]);
}
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");
printf("arg4 to 17: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
"BatchStrideB0, BatchStrideB1, BatchStrideC\n");
exit(0);
}
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? O : M;
StrideA = (StrideA < 0) ? DefaultStrideA : StrideA;
StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
StrideC = (StrideC < 0) ? DefaultStrideC : StrideC;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
BatchStrideA = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
BatchStrideC = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
auto f_host_tensor_descriptor = [](std::size_t batch_count,
std::size_t row,
std::size_t col,
std::size_t stride,
std::size_t batch_stride,
auto layout) {
if(std::is_same<decltype(layout), Row>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
std::vector<std::size_t>({batch_stride, 1, stride}));
}
};
// C_m_o = A_m_k * B0_k_n * B1_n_o
Tensor<ADataType> a_g_m_k(
f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
Tensor<B0DataType> b0_g_k_n(
f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
Tensor<B1DataType> b1_g_n_o(
f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
Tensor<CDataType> c_g_m_o_host_result(
f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
Tensor<CDataType> c_g_m_o_device_result(
f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
break;
case 2:
a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
break;
default:
a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
}
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSize());
DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
DeviceMem c_g_m_o_device_buf(sizeof(CDataType) * c_g_m_o_device_result.mDesc.GetElementSize());
a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument =
gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
M,
N,
K,
O,
BatchCount,
StrideA,
StrideB0,
StrideB1,
StrideC,
BatchStrideA,
BatchStrideB0,
BatchStrideB1,
BatchStrideC,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
BatchCount;
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, "
<< gemm.GetTypeString() << std::endl;
c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
if(do_verification)
{
// Output of Gemm0 is input A of Gemm1
Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, a1_g_m_n, a_element_op, b0_element_op, PassThrough{});
ref_gemm0_invoker.Run(ref_gemm0_argument);
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(
a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
return ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData) ? 0 : 1;
}
return 0;
}

0
example/32_batched_gemm_gemm/CMakeLists.txt → example/32_batched_gemm_softmax_gemm/CMakeLists.txt
View File

0
example/32_batched_gemm_gemm/batched_gemm_softmax_gemm_xdl_fp16.cpp → example/32_batched_gemm_softmax_gemm/batched_gemm_softmax_gemm_xdl_fp16.cpp
View File

6
example/CMakeLists.txt
View File

@@ -44,6 +44,6 @@ add_subdirectory(26_contraction)
add_subdirectory(27_layernorm)
add_subdirectory(28_grouped_gemm_bias_e_permute)
add_subdirectory(29_batched_gemm_bias_e_permute)
add_subdirectory(30_grouped_convnd_fwd_bias_relu)
add_subdirectory(31_grouped_convnd_fwd_bias_relu_add)
add_subdirectory(32_batched_gemm_gemm)
add_subdirectory(30_grouped_convnd_fwd_bias_relu_add)
add_subdirectory(31_batched_gemm_gemm)
add_subdirectory(32_batched_gemm_softmax_gemm)