Skip to content

GitLab

Commit 4ed59413 authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into fix_0813

parents 8bea6b2d 0bd6b842
Show whitespace changes
Inline Side-by-side
Showing with 1905 additions and 681 deletions
example/27_layernorm/layernorm_blockwise.cpp
View file @ 4ed59413
...@@ -9,7 +9,7 @@ ...@@ -9,7 +9,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp" #include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/device_layernorm.hpp" #include "ck/tensor_operation/gpu/device/device_layernorm_impl.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp" #include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -29,7 +29,7 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough; ...@@ -29,7 +29,7 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
constexpr int Rank = 2; constexpr int Rank = 2;
constexpr int NumReduceDim = 1; constexpr int NumReduceDim = 1;
using DeviceInstance = ck::tensor_operation::device::DeviceLayernorm<XDataType, using DeviceInstance = ck::tensor_operation::device::DeviceLayernormImpl<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
AccDataType, AccDataType,
...@@ -90,6 +90,7 @@ int main() ...@@ -90,6 +90,7 @@ int main()
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()}, std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
std::vector<ck::index_t>{gamma.mDesc.GetStrides().begin(), gamma.mDesc.GetStrides().end()}, std::vector<ck::index_t>{gamma.mDesc.GetStrides().begin(), gamma.mDesc.GetStrides().end()},
std::vector<ck::index_t>{beta.mDesc.GetStrides().begin(), beta.mDesc.GetStrides().end()}, std::vector<ck::index_t>{beta.mDesc.GetStrides().begin(), beta.mDesc.GetStrides().end()},
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
{1}, {1},
1e-4, 1e-4,
x_dev.GetDeviceBuffer(), x_dev.GetDeviceBuffer(),
......
example/30_grouped_convnd_fwd_bias_relu/CMakeLists.txt deleted 100644 → 0
View file @ 8bea6b2d
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)
example/30_grouped_convnd_fwd_bias_relu/README.md deleted 100644 → 0
View file @ 8bea6b2d
```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>
```
example/30_grouped_convnd_fwd_bias_relu/grouped_convnd_fwd_bias_common.hpp deleted 100644 → 0
View file @ 8bea6b2d
// 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;
}
example/30_grouped_convnd_fwd_bias_relu/grouped_convnd_fwd_bias_relu_xdl_fp16.cpp deleted 100644 → 0
View file @ 8bea6b2d
// 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;
}
example/31_batched_gemm_gemm/CMakeLists.txt 0 → 100644
View file @ 4ed59413
add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp 0 → 100644
View file @ 4ed59413
// 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;
}
example/CMakeLists.txt
View file @ 4ed59413
...@@ -44,6 +44,6 @@ add_subdirectory(26_contraction) ...@@ -44,6 +44,6 @@ add_subdirectory(26_contraction)
add_subdirectory(27_layernorm) add_subdirectory(27_layernorm)
add_subdirectory(28_grouped_gemm_bias_e_permute) add_subdirectory(28_grouped_gemm_bias_e_permute)
add_subdirectory(29_batched_gemm_bias_e_permute) add_subdirectory(29_batched_gemm_bias_e_permute)
add_subdirectory(30_grouped_convnd_fwd_bias_relu) add_subdirectory(30_grouped_convnd_fwd_bias_relu_add)
add_subdirectory(31_grouped_convnd_fwd_bias_relu_add) add_subdirectory(31_batched_gemm_gemm)
add_subdirectory(32_batched_gemm_gemm) add_subdirectory(32_batched_gemm_softmax_gemm)
include/ck/tensor_operation/gpu/block/blockwise_welford.hpp 0 → 100644
View file @ 4ed59413
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_description/cluster_descriptor.hpp"
#include "ck/utility/reduction_common.hpp"
namespace ck {
// clang-format off
// Assume:
// 1) work_buffer is buffer (typically LDS) allocated outside as workspace
// 2) work_buffer has T elements, and space size is no less than 3*BlockSize
// 3) mean_value, var_value and count is the input data in vgpr from each thread
// 4) mean_value, var_value and count is the over-written reduced output in vgpr for each thread
// 5) Merge mean and M from ThreadwiseWelford
// clang-format on
template <typename T,
index_t BlockSize,
typename ThreadClusterLengths_M_K,
typename ThreadClusterArrangeOrder,
bool GetActualVariance = true>
struct BlockwiseWelford
{
static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
"The product of cluster lengths should be same as BlockSize!");
static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
static constexpr auto thread_cluster_desc =
make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
__device__ static inline void
Merge(T& mean_a, T& var_a, int& count_a, T mean_b, T var_b, int count_b)
{
int count = count_a + count_b;
T count_b_over_count = count == 0 ? type_convert<T>(0) : type_convert<T>(count_b) / count;
T delta = mean_b - mean_a;
mean_a += delta * count_b_over_count;
var_a += var_b + delta * delta * count_a * count_b_over_count;
count_a = count;
}
__device__ static void Run(T& mean_value, T& var_value, int& count)
{
__shared__ T mean_block_buf[BlockSize];
__shared__ T var_block_buf[BlockSize];
__shared__ int count_block_buf[BlockSize];
constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
const auto thread_cluster_idx =
thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
mean_block_buf[offset1] = mean_value;
var_block_buf[offset1] = var_value;
count_block_buf[offset1] = count;
block_sync_lds();
static_for<0, cluster_len_shift, 1>{}([&](auto I) {
constexpr index_t indOffset = 1 << (cluster_len_shift - 1 - I());
if(thread_k_cluster_id < indOffset)
{
index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
make_tuple(0, indOffset));
T mean1 = mean_block_buf[offset1];
T var1 = var_block_buf[offset1];
int count1 = count_block_buf[offset1];
T mean2 = mean_block_buf[offset2];
T var2 = var_block_buf[offset2];
int count2 = count_block_buf[offset2];
Merge(mean1, var1, count1, mean2, var2, count2);
mean_block_buf[offset1] = mean1;
var_block_buf[offset1] = var1;
count_block_buf[offset1] = count1;
}
block_sync_lds();
});
index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
count = count_block_buf[offset];
mean_value = mean_block_buf[offset];
if constexpr(GetActualVariance)
var_value = var_block_buf[offset] / count;
else
var_value = var_block_buf[offset];
};
};
} // namespace ck
include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm_xdl_cshuffle.hpp 0 → 100644
View file @ 4ed59413
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename AccElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename B1GridDesc_BK0_N_BK1,
typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename Block2CTileMap,
typename ComputeBasePtrOfStridedBatch,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_gemm_xdl_cshuffle_v1(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
const FloatAB* __restrict__ p_b1_grid,
FloatC* __restrict__ p_c_grid,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const AccElementwiseOperation acc_element_op,
const B1ElementwiseOperation b1_element_op,
const CElementwiseOperation c_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2CTileMap block_2_ctile_map,
const index_t batch_count,
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetABasePtr(g_idx)));
const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(g_idx)));
const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(g_idx)));
const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset,
p_b1_grid + b1_batch_offset,
p_c_grid + c_batch_offset,
p_shared,
a_element_op,
b_element_op,
acc_element_op,
b1_element_op,
c_element_op,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
b1_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mperblock_nblock_nperblock,
block_2_ctile_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_b1_grid;
ignore = p_c_grid;
ignore = a_element_op;
ignore = b_element_op;
ignore = acc_element_op;
ignore = b1_element_op;
ignore = c_element_op;
ignore = a_grid_desc_ak0_m_ak1;
ignore = b_grid_desc_bk0_n_bk1;
ignore = b1_grid_desc_bk0_n_bk1;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = block_2_ctile_map;
ignore = batch_count;
ignore = compute_base_ptr_of_batch;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
// Computes C = A * B0 * B1
// ^^^^^^ (Acc0)
// ^^^^^^^^^^^ (Acc1)
template <typename ALayout,
typename BLayout, // B0Layout
typename B1Layout,
typename CLayout,
typename ADataType,
typename BDataType,
typename B1DataType,
typename CDataType,
typename GemmAccDataType,
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename AccElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation,
GemmSpecialization GemmSpec,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock, // Gemm0NPerBlock
index_t KPerBlock, // Gemm0KPerBlock
index_t Gemm1NPerBlock,
index_t Gemm1KPerBlock,
index_t AK1,
index_t BK1,
index_t B1K1,
index_t MPerXDL,
index_t NPerXDL,
index_t MXdlPerWave,
index_t NXdlPerWave,
index_t Gemm1NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
typename B1BlockTransferThreadClusterArrangeOrder,
typename B1BlockTransferSrcAccessOrder,
index_t B1BlockTransferSrcVectorDim,
index_t B1BlockTransferSrcScalarPerVector,
index_t B1BlockTransferDstScalarPerVector_BK1,
bool B1BlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
LoopScheduler LoopSched = LoopScheduler::Default>
struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout,
BLayout,
B1Layout,
CLayout,
ADataType,
BDataType,
B1DataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation>
{
using DeviceOp = DeviceBatchedGemmGemm_Xdl_CShuffle;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
{
const auto a_grid_desc_mraw_kraw = [&]() {
if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(StrideA, I1));
}
else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
make_tuple(I1, StrideA));
}
}();
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
const auto MPad = M - MRaw;
const auto KPad = K - KRaw;
if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both M and K
assert(K % AK1 == 0);
const auto AK0 = K / AK1;
const auto a_grid_desc_m_k =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad M, but not K
assert(KRaw % AK1 == 0);
const auto AK0 = KRaw / AK1;
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_right_pad_transform(MRaw, MPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad K, but not M
assert(K % AK1 == 0);
const auto AK0 = K / AK1;
const auto a_grid_desc_m_k = transform_tensor_descriptor(
a_grid_desc_mraw_kraw,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(MRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
else
{
// not pad M or K
assert(KRaw % AK1 == 0);
const auto AK0 = KRaw / AK1;
const auto a_grid_desc_ak0_m_ak1 =
transform_tensor_descriptor(a_grid_desc_mraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(MRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return a_grid_desc_ak0_m_ak1;
}
}
static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
{
const auto b_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(I1, StrideB));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(StrideB, I1));
}
}();
const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
const auto NPad = N - NRaw;
const auto KPad = K - KRaw;
if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad both N and K
const auto BK0 = K / BK1;
const auto b_grid_desc_n_k =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_right_pad_transform(NRaw, NPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::MNPadding)
{
// pad N, but not K
const auto BK0 = KRaw / BK1;
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad K, but not N
const auto BK0 = K / BK1;
const auto b_grid_desc_n_k = transform_tensor_descriptor(
b_grid_desc_nraw_kraw,
make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
else
{
// not pad N or K
const auto BK0 = KRaw / BK1;
const auto b_grid_desc_bk0_n_bk1 =
transform_tensor_descriptor(b_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b_grid_desc_bk0_n_bk1;
}
}
// Args: Gemm1KRaw, Gemm1NRaw, StrideB1
static auto MakeB1GridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
{
const auto b1_grid_desc_nraw_kraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, B1Layout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(I1, StrideB));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B1Layout>::value)
{
return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
make_tuple(StrideB, I1));
}
}();
const auto N = math::integer_divide_ceil(NRaw, Gemm1NPerBlock) * Gemm1NPerBlock;
const auto K = math::integer_divide_ceil(KRaw, Gemm1KPerBlock) * Gemm1KPerBlock;
const auto NPad = N - NRaw;
const auto KPad = K - KRaw;
// TODO: implement finer-grained padding
if constexpr(GemmSpec == GemmSpecialization::Default)
{
const auto B1K0 = KRaw / B1K1;
const auto b1_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b1_grid_desc_nraw_kraw,
make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b1_grid_desc_bk0_n_bk1;
}
else
{
// pad both B1N and B1K
const auto B1K0 = K / B1K1;
const auto b1_grid_desc_n_k =
transform_tensor_descriptor(b1_grid_desc_nraw_kraw,
make_tuple(make_right_pad_transform(NRaw, NPad),
make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
const auto b1_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
b1_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
return b1_grid_desc_bk0_n_bk1;
}
}
static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
{
const auto c_grid_desc_mraw_nraw = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
make_tuple(I1, StrideC));
}
}();
const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
const auto N = math::integer_divide_ceil(NRaw, Gemm1NPerBlock) * Gemm1NPerBlock;
const auto MPad = M - MRaw;
const auto NPad = N - NRaw;
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(MRaw, MPad),
make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad M, but not N
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad N, but not M
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
// not pad M or N
return c_grid_desc_mraw_nraw;
}
}
struct ComputeBasePtrOfStridedBatch
{
ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
index_t BatchStrideB,
index_t BatchStrideB1,
index_t BatchStrideC)
: BatchStrideA_(BatchStrideA),
BatchStrideB_(BatchStrideB),
BatchStrideB1_(BatchStrideB1),
BatchStrideC_(BatchStrideC)
{
}
__host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideA_);
}
__host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideB_);
}
__host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideB1_);
}
__host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
{
return g_idx * static_cast<long_index_t>(BatchStrideC_);
}
private:
index_t BatchStrideA_;
index_t BatchStrideB_;
index_t BatchStrideB1_;
index_t BatchStrideC_;
};
using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1(1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
// GridwiseGemm
using GridwiseGemm = GridwiseBatchedGemmGemm_Xdl_CShuffle<
ADataType, // TODO: distinguish A/B datatype
GemmAccDataType,
CShuffleDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation,
InMemoryDataOperationEnum::Set,
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
B1GridDesc_BK0_N_BK1,
CGridDesc_M_N,
NumGemmKPrefetchStage,
BlockSize,
MPerBlock,
NPerBlock,
KPerBlock,
Gemm1NPerBlock,
Gemm1KPerBlock,
AK1,
BK1,
B1K1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
Gemm1NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
true,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
true,
BBlockLdsExtraN,
B1BlockTransferThreadClusterLengths_BK0_N_BK1,
B1BlockTransferThreadClusterArrangeOrder,
B1BlockTransferSrcAccessOrder,
B1BlockTransferSrcVectorDim,
B1BlockTransferSrcScalarPerVector,
B1BlockTransferDstScalarPerVector_BK1,
false,
B1BlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CShuffleBlockTransferScalarPerVector_NPerBlock,
LoopSched>;
// Argument
struct Argument : public BaseArgument
{
Argument(const ADataType* p_a_grid,
const BDataType* p_b_grid,
const B1DataType* p_b1_grid,
CDataType* p_c_grid,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t Gemm1NRaw, // = ORaw
index_t Batch,
index_t StrideA,
index_t StrideB,
index_t StrideB1,
index_t StrideC,
index_t BatchStrideA,
index_t BatchStrideB,
index_t BatchStrideB1,
index_t BatchStrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
AccElementwiseOperation acc_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_b1_grid_{p_b1_grid},
p_c_grid_{p_c_grid},
a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
b1_grid_desc_bk0_n_bk1_{
DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(NRaw, Gemm1NRaw, StrideB1)},
c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, Gemm1NRaw, StrideC)},
c_grid_desc_mblock_mperblock_nblock_nperblock_{},
block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
acc_element_op_{acc_element_op},
b1_element_op_{b1_element_op},
c_element_op_{c_element_op},
batch_count_(Batch),
compute_base_ptr_of_batch_{BatchStrideA, BatchStrideB, BatchStrideB1, BatchStrideC}
{
if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
b_grid_desc_bk0_n_bk1_,
b1_grid_desc_bk0_n_bk1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n_);
}
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
const B1DataType* p_b1_grid_;
CDataType* p_c_grid_;
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
CGridDesc_M_N c_grid_desc_m_n_;
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock_;
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
AccElementwiseOperation acc_element_op_;
B1ElementwiseOperation b1_element_op_;
CElementwiseOperation c_element_op_;
index_t batch_count_;
ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
};
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceOp::Argument;
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.b1_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_))
{
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.batch_count_;
// Gemm0_K
const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
float ave_time = 0;
auto launch_kernel = [&](auto has_main_k_block_loop_) {
const auto kernel = kernel_gemm_gemm_xdl_cshuffle_v1<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
AElementwiseOperation,
BElementwiseOperation,
AccElementwiseOperation,
B1ElementwiseOperation,
CElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
DeviceOp::B1GridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename GridwiseGemm::DefaultBlock2CTileMap,
ComputeBasePtrOfStridedBatch,
has_main_k_block_loop_>;
return launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_b1_grid_,
arg.p_c_grid_,
arg.a_element_op_,
arg.b_element_op_,
arg.acc_element_op_,
arg.b1_element_op_,
arg.c_element_op_,
arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.b1_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.block_2_ctile_map_,
arg.batch_count_,
arg.compute_base_ptr_of_batch_);
};
// Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
// to concern Gemm0's loop
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
ave_time = launch_kernel(integral_constant<bool, true>{});
}
else
{
ave_time = launch_kernel(integral_constant<bool, false>{});
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
{
return false;
}
return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_,
arg.b1_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
const B1DataType* p_b1,
CDataType* p_c,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t Gemm1NRaw,
index_t Batch,
index_t StrideA,
index_t StrideB,
index_t StrideB1,
index_t StrideC,
index_t BatchStrideA,
index_t BatchStrideB,
index_t BatchStrideB1,
index_t BatchStrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
AccElementwiseOperation acc_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op)
{
return Argument{p_a, p_b, p_b1, p_c, MRaw,
NRaw, KRaw, Gemm1NRaw, Batch, StrideA,
StrideB, StrideB1, StrideC, BatchStrideA, BatchStrideB,
BatchStrideB1, BatchStrideC, a_element_op, b_element_op, acc_element_op,
b1_element_op, c_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
const void* p_b1,
void* p_c,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t Gemm1NRaw,
index_t Batch,
index_t StrideA,
index_t StrideB,
index_t StrideB1,
index_t StrideC,
index_t BatchStrideA,
index_t BatchStrideB,
index_t BatchStrideB1,
index_t BatchStrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
AccElementwiseOperation acc_element_op,
B1ElementwiseOperation b1_element_op,
CElementwiseOperation c_element_op) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<const B1DataType*>(p_b1),
static_cast<CDataType*>(p_c),
MRaw,
NRaw,
KRaw,
Gemm1NRaw,
Batch,
StrideA,
StrideB,
StrideB1,
StrideC,
BatchStrideA,
BatchStrideB,
BatchStrideB1,
BatchStrideC,
a_element_op,
b_element_op,
acc_element_op,
b1_element_op,
c_element_op);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "DeviceBatchedGemmGemm_Xdl_CShuffle"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
<< NPerBlock << ", "
<< KPerBlock << ", "
<< AK1 << ", "
<< BK1 << ", "
<< MPerBlock << ", "
<< Gemm1NPerBlock << ", "
<< Gemm1KPerBlock << ", "
<< B1K1 << ">";
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_layernorm.hpp → include/ck/tensor_operation/gpu/device/device_layernorm_impl.hpp
View file @ 4ed59413
...@@ -9,13 +9,48 @@ ...@@ -9,13 +9,48 @@
#include "ck/utility/reduction_operator.hpp" #include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/device_normalization.hpp" #include "ck/tensor_operation/gpu/device/device_normalization.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp" #include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_multiblock.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_common.hpp" #include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_layernorm.hpp" #include "ck/tensor_operation/gpu/grid/gridwise_layernorm_welford_variance.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp" #include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
#include "ck/host_utility/device_prop.hpp" #include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp" #include "ck/host_utility/kernel_launch.hpp"
namespace ck {
template <typename GridwiseReduction,
typename XDataType,
typename GammaDataType,
typename BetaDataType,
typename YDataType,
typename AccDataType,
typename AccElementwiseOperation,
typename GridDesc_M_K,
typename GridDesc_K>
__global__ void kernel_layernorm(const GridDesc_M_K x_grid_desc_m_k,
const GridDesc_K gamma_grid_desc_k,
const GridDesc_K beta_grid_desc_k,
const GridDesc_M_K y_grid_desc_m_k,
index_t num_k_block_tile_iteration,
AccDataType epsilon,
const XDataType* const __restrict__ p_x_global,
const GammaDataType* const __restrict__ p_gamma_global,
const BetaDataType* const __restrict__ p_beta_global,
YDataType* const __restrict__ p_y_global,
const AccElementwiseOperation acc_elementwise_op)
{
GridwiseReduction::Run(x_grid_desc_m_k,
gamma_grid_desc_k,
beta_grid_desc_k,
y_grid_desc_m_k,
num_k_block_tile_iteration,
epsilon,
p_x_global,
p_gamma_global,
p_beta_global,
p_y_global,
acc_elementwise_op);
};
} // namespace ck
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
...@@ -39,7 +74,7 @@ template <typename XDataType, ...@@ -39,7 +74,7 @@ template <typename XDataType,
index_t GammaSrcVectorSize, index_t GammaSrcVectorSize,
index_t BetaSrcVectorSize, index_t BetaSrcVectorSize,
index_t YDstVectorSize> index_t YDstVectorSize>
struct DeviceLayernorm : public DeviceNormalization2<XDataType, struct DeviceLayernormImpl : public DeviceLayernorm<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
AccDataType, AccDataType,
...@@ -58,27 +93,74 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType, ...@@ -58,27 +93,74 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType,
using PassThrough = tensor_operation::element_wise::PassThrough; using PassThrough = tensor_operation::element_wise::PassThrough;
// Used for freeloading of some handy functions from DeviceReduceMultiBlock static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
using Reduction = DeviceReduceMultiBlock<XDataType, static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
AccDataType,
YDataType, static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
Rank, const std::vector<index_t>& inStrides,
NumReduceDim, int blkGroupSize,
reduce::Add, int numBlockTileIteration)
PassThrough, // InElementwiseOperation {
AccElementwiseOperation, // AccElementwiseOperation constexpr index_t NumInvariantDim = Rank - NumReduceDim;
InMemoryDataOperationEnum::Set, static constexpr index_t numSrcDim = Rank;
false, // PropagateNan static constexpr bool reduceAllDim = (NumInvariantDim == 0);
false, // OutputIndex
false, // HaveIndexInputIfOutputIndex const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
BlockSize, const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
MThreadClusterSize,
KThreadClusterSize, const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
MThreadSliceSize,
KThreadSliceSize, const auto in_grid_desc_m_k = [&]() {
XYSrcVectorDim, if constexpr(reduceAllDim)
XSrcVectorSize, {
1>; // YDstVectorSize const auto one_dim_inDesc = transform_tensor_descriptor(
inDesc,
make_tuple(make_merge_transform(tupleSrcLengths)),
make_tuple(typename arithmetic_sequence_gen<0, numSrcDim, 1>::type{}),
make_tuple(Sequence<0>{}));
return transform_tensor_descriptor(one_dim_inDesc,
make_tuple(make_unmerge_transform(make_tuple(
1, one_dim_inDesc.GetLength(Number<0>{})))),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0, 1>{}));
}
else
{
using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
const auto reduceDimLengths =
make_tuple_from_array_and_index_seq(inLengths, ReduceDims{});
const auto invariantDimLengths =
make_tuple_from_array_and_index_seq(inLengths, InvariantDims{});
return transform_tensor_descriptor(
inDesc,
make_tuple(make_merge_transform(invariantDimLengths),
make_merge_transform(reduceDimLengths)),
make_tuple(InvariantDims{}, ReduceDims{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
}();
const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
const auto reduceLength = in_grid_desc_m_k.GetLength(Number<1>{});
const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
const auto inPad_M =
math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
in_grid_desc_m_k,
make_tuple(make_right_pad_transform(invariantLength, inPad_M),
make_right_pad_transform(reduceLength, inPad_K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return (in_grid_desc_m_k_padded);
};
static auto MakeAffine1dDescriptor(const std::vector<index_t>& Lengths, static auto MakeAffine1dDescriptor(const std::vector<index_t>& Lengths,
const std::vector<index_t>& Strides, const std::vector<index_t>& Strides,
...@@ -97,7 +179,7 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType, ...@@ -97,7 +179,7 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType,
make_tuple(Sequence<0>{})); make_tuple(Sequence<0>{}));
const auto reduceTotalLength = grid_desc_k.GetLength(Number<0>{}); const auto reduceTotalLength = grid_desc_k.GetLength(Number<0>{});
const int reduceSizePerBlock = Reduction::K_BlockTileSize * numBlockTileIteration; const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
const auto Pad_K = reduceSizePerBlock * blkGroupSize - reduceTotalLength; const auto Pad_K = reduceSizePerBlock * blkGroupSize - reduceTotalLength;
...@@ -110,10 +192,11 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType, ...@@ -110,10 +192,11 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType,
return (grid_desc_k_padded); return (grid_desc_k_padded);
}; };
using GridDesc_M_K = decltype(Reduction::MakeSrc2dDescriptor({1}, {1}, 1, 1)); using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
using GridDesc_K = decltype(MakeAffine1dDescriptor({1}, {1}, 1, 1)); using GridDesc_K = decltype(MakeAffine1dDescriptor({1}, {1}, 1, 1));
using GridwiseReduceLayernormGeneric = GridwiseLayernorm_mk_to_mk<XDataType, using GridwiseReduceLayernormGeneric =
GridwiseLayernormWelfordVariance_mk_to_mk<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
...@@ -134,7 +217,8 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType, ...@@ -134,7 +217,8 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType,
YDstVectorSize, YDstVectorSize,
false>; false>;
using GridwiseReduceLayernormSweepOnce = GridwiseLayernorm_mk_to_mk<XDataType, using GridwiseReduceLayernormSweepOnce =
GridwiseLayernormWelfordVariance_mk_to_mk<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
YDataType, YDataType,
...@@ -155,12 +239,13 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType, ...@@ -155,12 +239,13 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType,
YDstVectorSize, YDstVectorSize,
true>; true>;
struct Argument : public Reduction::Argument struct Argument : public BaseArgument
{ {
Argument(const std::vector<index_t> lengths, Argument(const std::vector<index_t> lengths,
const std::vector<index_t> xStrides, const std::vector<index_t> xStrides,
const std::vector<index_t> gammaStrides, const std::vector<index_t> gammaStrides,
const std::vector<index_t> betaStrides, const std::vector<index_t> betaStrides,
const std::vector<index_t> yStrides,
const std::vector<index_t> reduceDims, const std::vector<index_t> reduceDims,
AccElementwiseOperation acc_elementwise_op, AccElementwiseOperation acc_elementwise_op,
AccDataType epsilon, AccDataType epsilon,
...@@ -168,53 +253,76 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType, ...@@ -168,53 +253,76 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType,
const GammaDataType* p_gamma, const GammaDataType* p_gamma,
const BetaDataType* p_beta, const BetaDataType* p_beta,
YDataType* p_y) YDataType* p_y)
: Reduction::Argument(lengths, : epsilon_(epsilon),
xStrides, p_x_(p_x),
{},
{},
reduceDims,
0.0f, // alpha
0.0f, // beta
p_x,
nullptr,
p_y,
nullptr,
acc_elementwise_op,
PassThrough{}),
epsilon_(epsilon),
p_gamma_(p_gamma), p_gamma_(p_gamma),
p_beta_(p_beta), p_beta_(p_beta),
p_y_(p_y),
gammaStrides_(gammaStrides), gammaStrides_(gammaStrides),
betaStrides_(betaStrides) betaStrides_(betaStrides),
acc_elementwise_op_(acc_elementwise_op)
{ {
reduceLength_.resize(NumReduceDim); Lengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(lengths, reduceDims);
xStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(xStrides, reduceDims);
yStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
long_index_t invariant_total_length;
long_index_t reduce_total_length;
std::tie(invariant_total_length, reduce_total_length) =
get_2d_lengths<Rank, NumReduceDim>(Lengths_);
blkGroupSize_ = 1;
numBlockTileIteration_ = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
gridSize_ = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
M_BlockTileSize * blkGroupSize_;
reduceLengths_.resize(NumReduceDim);
for(int i = 0; i < NumReduceDim; ++i) for(int i = 0; i < NumReduceDim; ++i)
{ {
reduceLength_[i] = lengths[reduceDims[i]]; reduceLengths_[i] = lengths[reduceDims[i]];
} }
} }
AccDataType epsilon_; AccDataType epsilon_;
const XDataType* p_x_;
const GammaDataType* p_gamma_; const GammaDataType* p_gamma_;
const BetaDataType* p_beta_; const BetaDataType* p_beta_;
std::vector<index_t> reduceLength_; YDataType* p_y_;
std::vector<index_t> Lengths_;
std::vector<index_t> xStrides_;
std::vector<index_t> reduceLengths_;
std::vector<index_t> gammaStrides_; std::vector<index_t> gammaStrides_;
std::vector<index_t> betaStrides_; std::vector<index_t> betaStrides_;
std::vector<index_t> yStrides_;
AccElementwiseOperation acc_elementwise_op_;
int blkGroupSize_;
int numBlockTileIteration_;
size_t gridSize_;
}; };
struct Invoker : public BaseInvoker struct Invoker : public BaseInvoker
{ {
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{}) float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{ {
const auto x_grid_desc_m_k = Reduction::MakeSrc2dDescriptor( const auto x_grid_desc_m_k = MakeSrc2dDescriptor(
arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.numBlockTileIteration); arg.Lengths_, arg.xStrides_, arg.blkGroupSize_, arg.numBlockTileIteration_);
const auto gamma_grid_desc_k = MakeAffine1dDescriptor( const auto gamma_grid_desc_k = MakeAffine1dDescriptor(arg.reduceLengths_,
arg.reduceLength_, arg.gammaStrides_, arg.blkGroupSize, arg.numBlockTileIteration); arg.gammaStrides_,
const auto beta_grid_desc_k = MakeAffine1dDescriptor( arg.blkGroupSize_,
arg.reduceLength_, arg.betaStrides_, arg.blkGroupSize, arg.numBlockTileIteration); arg.numBlockTileIteration_);
const auto y_grid_desc_m_k = Reduction::MakeSrc2dDescriptor( const auto beta_grid_desc_k = MakeAffine1dDescriptor(arg.reduceLengths_,
arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.numBlockTileIteration); arg.betaStrides_,
arg.blkGroupSize_,
arg.numBlockTileIteration_);
const auto y_grid_desc_m_k = MakeSrc2dDescriptor(
arg.Lengths_, arg.yStrides_, arg.blkGroupSize_, arg.numBlockTileIteration_);
bool sweep_once = bool sweep_once =
x_grid_desc_m_k.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize; x_grid_desc_m_k.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
...@@ -241,19 +349,19 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType, ...@@ -241,19 +349,19 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType,
float avg_time = 0; float avg_time = 0;
avg_time += launch_and_time_kernel(stream_config, avg_time += launch_and_time_kernel(stream_config,
kernel_main, kernel_main,
dim3(arg.gridSize), dim3(arg.gridSize_),
dim3(BlockSize), dim3(BlockSize),
0, 0,
x_grid_desc_m_k, x_grid_desc_m_k,
gamma_grid_desc_k, gamma_grid_desc_k,
beta_grid_desc_k, beta_grid_desc_k,
y_grid_desc_m_k, y_grid_desc_m_k,
arg.numBlockTileIteration, arg.numBlockTileIteration_,
arg.epsilon_, arg.epsilon_,
arg.in_dev_, arg.p_x_,
arg.p_gamma_, arg.p_gamma_,
arg.p_beta_, arg.p_beta_,
arg.out_dev_, arg.p_y_,
arg.acc_elementwise_op_); arg.acc_elementwise_op_);
return (avg_time); return (avg_time);
...@@ -270,12 +378,33 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType, ...@@ -270,12 +378,33 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType,
{ {
const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg); const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
if(!Reduction::IsSupportedArgument(p_arg_)) constexpr index_t NumInvariantDim = Rank - NumReduceDim;
if constexpr(XYSrcVectorDim == 0)
{
if constexpr(NumInvariantDim == 0)
{
return false;
}
else
{ {
if(p_arg_->xStrides_[NumInvariantDim - 1] != 1)
return false; return false;
if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
return false;
};
} }
else
{
if(p_arg_->xStrides_[Rank - 1] != 1)
return false;
if(p_arg_->Lengths_[Rank - 1] % XSrcVectorSize != 0)
return false;
};
if(p_arg_->inLengths_[Rank - 1] % YDstVectorSize != 0) if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0)
{ {
return false; return false;
} }
...@@ -309,6 +438,7 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType, ...@@ -309,6 +438,7 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType,
const std::vector<index_t> xStrides, const std::vector<index_t> xStrides,
const std::vector<index_t> gammaStrides, const std::vector<index_t> gammaStrides,
const std::vector<index_t> betaStrides, const std::vector<index_t> betaStrides,
const std::vector<index_t> yStrides,
const std::vector<index_t> reduceDims, const std::vector<index_t> reduceDims,
AccDataType epsilon, AccDataType epsilon,
const void* p_x, const void* p_x,
...@@ -321,6 +451,7 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType, ...@@ -321,6 +451,7 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType,
xStrides, xStrides,
gammaStrides, gammaStrides,
betaStrides, betaStrides,
yStrides,
reduceDims, reduceDims,
acc_elementwise_op, acc_elementwise_op,
epsilon, epsilon,
...@@ -340,7 +471,7 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType, ...@@ -340,7 +471,7 @@ struct DeviceLayernorm : public DeviceNormalization2<XDataType,
auto str = std::stringstream(); auto str = std::stringstream();
// clang-format off // clang-format off
str << "DeviceLayernorm<" << BlockSize << ","; str << "DeviceLayernormImpl<" << BlockSize << ",";
str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ","; str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ","; str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
str << "XYSrcVectorDim_" << XYSrcVectorDim << ","; str << "XYSrcVectorDim_" << XYSrcVectorDim << ",";
......
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