Skip to content

GitLab

Commit f0224f2a authored by letaoqin's avatar letaoqin
Browse files

Merge branch 'develop' into dl_conv_multiple_d

parents befc2638 0e9c88ce
Show whitespace changes
Inline Side-by-side
Showing with 550 additions and 440 deletions
example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
View file @ f0224f2a
...@@ -160,14 +160,12 @@ bool run_gemm_reduce_add_addsquare_xdl(ck::index_t M, ...@@ -160,14 +160,12 @@ bool run_gemm_reduce_add_addsquare_xdl(ck::index_t M,
{ {
case 0: break; case 0: break;
case 1: case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k.begin(), ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
a_m_k.end()); ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n.begin(),
b_k_n.end());
break; break;
default: default:
ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k.begin(), a_m_k.end()); ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n.begin(), b_k_n.end()); ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
break; break;
} }
...@@ -264,15 +262,13 @@ bool run_gemm_reduce_add_addsquare_xdl(ck::index_t M, ...@@ -264,15 +262,13 @@ bool run_gemm_reduce_add_addsquare_xdl(ck::index_t M,
Tensor<EDataType> e_m_n_host_converted(e_m_n_host); Tensor<EDataType> e_m_n_host_converted(e_m_n_host);
pass = ck::utils::check_err( pass = ck::utils::check_err(
e_m_n.mData, e_m_n_host_converted.mData, "Error: Incorrect results c", 1e-2, 1e-2); e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
r0_device_buf.FromDevice(r0_m.mData.data()); r0_device_buf.FromDevice(r0_m.mData.data());
r1_device_buf.FromDevice(r1_m.mData.data()); r1_device_buf.FromDevice(r1_m.mData.data());
pass &= ck::utils::check_err( pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
r0_m.mData, r0_m_host.mData, "Error: Incorrect results d0", 1e-2, 1e-2); pass &= ck::utils::check_err(r1_m, r1_m_host, "Error: Incorrect results d1", 1e-2, 1e-2);
pass &= ck::utils::check_err(
r1_m.mData, r1_m_host.mData, "Error: Incorrect results d1", 1e-2, 1e-2);
if(pass) if(pass)
{ {
......
example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
View file @ f0224f2a
...@@ -134,14 +134,12 @@ auto run_gemm_reduce_max_xdl(ck::index_t M, ...@@ -134,14 +134,12 @@ auto run_gemm_reduce_max_xdl(ck::index_t M,
{ {
case 0: break; case 0: break;
case 1: case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k.begin(), ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
a_m_k.end()); ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n.begin(),
b_k_n.end());
break; break;
default: default:
ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k.begin(), a_m_k.end()); ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n.begin(), b_k_n.end()); ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
break; break;
} }
...@@ -243,8 +241,8 @@ auto run_gemm_reduce_max_xdl(ck::index_t M, ...@@ -243,8 +241,8 @@ auto run_gemm_reduce_max_xdl(ck::index_t M,
if constexpr(std::is_same_v<ADataType, ck::int4_t>) if constexpr(std::is_same_v<ADataType, ck::int4_t>)
{ {
Tensor<EDataType> e_m_n_device_converted(e_m_n); Tensor<EDataType> e_m_n_device_converted(e_m_n);
pass = ck::utils::check_err(e_m_n_device_converted.mData, pass = ck::utils::check_err(e_m_n_device_converted,
e_m_n_host_converted.mData, e_m_n_host_converted,
"Error: Incorrect results c", "Error: Incorrect results c",
1e-2, 1e-2,
1e-2); 1e-2);
...@@ -253,12 +251,11 @@ auto run_gemm_reduce_max_xdl(ck::index_t M, ...@@ -253,12 +251,11 @@ auto run_gemm_reduce_max_xdl(ck::index_t M,
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4 #endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
{ {
pass = ck::utils::check_err( pass = ck::utils::check_err(
e_m_n.mData, e_m_n_host_converted.mData, "Error: Incorrect results c", 1e-2, 1e-2); e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
} }
r0_device_buf.FromDevice(r0_m.mData.data()); r0_device_buf.FromDevice(r0_m.mData.data());
pass &= ck::utils::check_err( pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
r0_m.mData, r0_m_host.mData, "Error: Incorrect results d0", 1e-2, 1e-2);
if(pass) if(pass)
{ {
...@@ -339,14 +336,12 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M, ...@@ -339,14 +336,12 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M,
{ {
case 0: break; case 0: break;
case 1: case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k.begin(), ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
a_m_k.end()); ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n.begin(),
b_k_n.end());
break; break;
default: default:
ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k.begin(), a_m_k.end()); ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n.begin(), b_k_n.end()); ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
break; break;
} }
...@@ -460,8 +455,8 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M, ...@@ -460,8 +455,8 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M,
if constexpr(std::is_same_v<ADataType, ck::int4_t>) if constexpr(std::is_same_v<ADataType, ck::int4_t>)
{ {
Tensor<EDataType> e_m_n_device_converted(e_m_n); Tensor<EDataType> e_m_n_device_converted(e_m_n);
pass = ck::utils::check_err(e_m_n_device_converted.mData, pass = ck::utils::check_err(e_m_n_device_converted,
e_m_n_host_converted.mData, e_m_n_host_converted,
"Error: Incorrect results c", "Error: Incorrect results c",
1e-2, 1e-2,
1e-2); 1e-2);
...@@ -470,16 +465,14 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M, ...@@ -470,16 +465,14 @@ bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M,
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4 #endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
{ {
pass = ck::utils::check_err( pass = ck::utils::check_err(
e_m_n.mData, e_m_n_host_converted.mData, "Error: Incorrect results c", 1e-2, 1e-2); e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
} }
r0_device_buf.FromDevice(r0_m.mData.data()); r0_device_buf.FromDevice(r0_m.mData.data());
r1_device_buf.FromDevice(r1_m.mData.data()); r1_device_buf.FromDevice(r1_m.mData.data());
pass &= ck::utils::check_err( pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
r0_m.mData, r0_m_host.mData, "Error: Incorrect results d0", 1e-2, 1e-2); pass &= ck::utils::check_err(r1_m, r1_m_host, "Error: Incorrect results d1", 1e-2, 1e-2);
pass &= ck::utils::check_err(
r1_m.mData, r1_m_host.mData, "Error: Incorrect results d1", 1e-2, 1e-2);
if(pass) if(pass)
{ {
......
example/17_convnd_bwd_data/CMakeLists.txt
View file @ f0224f2a
add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp) add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp)
target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility) target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
View file @ f0224f2a
...@@ -61,9 +61,13 @@ int run_conv_bwd_data(bool do_verification, ...@@ -61,9 +61,13 @@ int run_conv_bwd_data(bool do_verification,
out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5}); out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5}); wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break; break;
default: case 2:
out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0}); out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5}); wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
break;
default:
out.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
} }
DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize()); DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
...@@ -98,9 +102,8 @@ int run_conv_bwd_data(bool do_verification, ...@@ -98,9 +102,8 @@ int run_conv_bwd_data(bool do_verification,
if(!conv.IsSupportedArgument(argument)) if(!conv.IsSupportedArgument(argument))
{ {
throw std::runtime_error( std::cout << "Not support,please check parameters or device";
"wrong! device_conv with the specified compilation parameters does " return 0;
"not support this Conv problem");
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
...@@ -142,7 +145,7 @@ int run_conv_bwd_data(bool do_verification, ...@@ -142,7 +145,7 @@ int run_conv_bwd_data(bool do_verification,
in_device_buf.FromDevice(in_device.mData.data()); in_device_buf.FromDevice(in_device.mData.data());
return ck::utils::check_err(in_device.mData, in_host.mData) ? 0 : 1; return ck::utils::check_err(in_device, in_host) ? 0 : 1;
} }
return 0; return 0;
......
example/20_convnd_bwd_weight/convnd_bwd_weight_xdl_fp16.cpp → example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
View file @ f0224f2a
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_bwd_weight_common.hpp" #include "convnd_bwd_data_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_weight_nwc_kxc_nwk_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp"
using InDataType = ck::half_t; using InDataType = ck::half_t;
using WeiDataType = ck::half_t; using WeiDataType = ck::half_t;
...@@ -17,61 +17,31 @@ using InElementOp = ck::tensor_operation::element_wise::PassThrough; ...@@ -17,61 +17,31 @@ using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough; using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough; using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdWeightDefault = static constexpr auto ConvBwdDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default; ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
template <ck::index_t NDimSpatial> template <ck::index_t NDimSpatial>
using DeviceConvndBwdWeightInstance = // clang-format off
ck::tensor_operation::device::DeviceConvNdBwdWeightNwcKxcNwk_Xdl_CShuffle< using DeviceConvNdBwdDataInstance = ck::tensor_operation::device::DeviceConvNdBwdDataNwcKxcNwk_Dl<
NDimSpatial, // NDimSpatial // ######| NDim| InData| WeiData| OutData| AccData| In| Wei| Out| Convolution| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
InDataType, // InDataType // ######| Spatial| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Forward| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
WeiDataType, // WeiDataType // ######| | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
OutDataType, // OutDataType // ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
AccDataType, // AccDataType NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType, InElementOp, WeiElementOp, OutElementOp, ConvBwdDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<1, 1, 8, 2>, S<16, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>;
InElementOp, // InElementwiseOperation // clang-format on
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
namespace ctc = ck::tensor_layout::convolution; namespace ctc = ck::tensor_layout::convolution;
print_helper_msg();
bool do_verification = true; bool do_verification = true;
int init_method = 1; int init_method = 1;
bool time_kernel = false; bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{ ck::utils::conv::ConvParam conv_param{
2, 1, 32, 256, 1024, {3, 3}, {14, 14}, {2, 2}, {1, 1}, {1, 1}, {1, 1}}; 2, 1, 128, 256, 256, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
ck::index_t split_k = 4;
if(argc == 1) if(argc == 1)
{ {
...@@ -91,9 +61,6 @@ int main(int argc, char* argv[]) ...@@ -91,9 +61,6 @@ int main(int argc, char* argv[])
const ck::index_t num_dim_spatial = std::stoi(argv[4]); const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv); conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
split_k = std::stoi(argv[5 + 3 + 6 * num_dim_spatial - 1]);
split_k = std::max(1, split_k);
} }
const auto in_element_op = InElementOp{}; const auto in_element_op = InElementOp{};
...@@ -118,14 +85,14 @@ int main(int argc, char* argv[]) ...@@ -118,14 +85,14 @@ int main(int argc, char* argv[])
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>( ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param); conv_param);
return run_conv_bwd_weight<1, return run_conv_bwd_data<1,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvndBwdWeightInstance<1>>(do_verification, DeviceConvNdBwdDataInstance<1>>(do_verification,
init_method, init_method,
time_kernel, time_kernel,
conv_param, conv_param,
...@@ -134,8 +101,7 @@ int main(int argc, char* argv[]) ...@@ -134,8 +101,7 @@ int main(int argc, char* argv[])
out_g_n_k_wos_desc, out_g_n_k_wos_desc,
in_element_op, in_element_op,
wei_element_op, wei_element_op,
out_element_op, out_element_op);
split_k);
} }
else if(conv_param.num_dim_spatial_ == 2) else if(conv_param.num_dim_spatial_ == 2)
{ {
...@@ -155,14 +121,14 @@ int main(int argc, char* argv[]) ...@@ -155,14 +121,14 @@ int main(int argc, char* argv[])
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>( ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param); conv_param);
return run_conv_bwd_weight<2, return run_conv_bwd_data<2,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvndBwdWeightInstance<2>>(do_verification, DeviceConvNdBwdDataInstance<2>>(do_verification,
init_method, init_method,
time_kernel, time_kernel,
conv_param, conv_param,
...@@ -171,8 +137,7 @@ int main(int argc, char* argv[]) ...@@ -171,8 +137,7 @@ int main(int argc, char* argv[])
out_g_n_k_wos_desc, out_g_n_k_wos_desc,
in_element_op, in_element_op,
wei_element_op, wei_element_op,
out_element_op, out_element_op);
split_k);
} }
else if(conv_param.num_dim_spatial_ == 3) else if(conv_param.num_dim_spatial_ == 3)
{ {
...@@ -192,14 +157,14 @@ int main(int argc, char* argv[]) ...@@ -192,14 +157,14 @@ int main(int argc, char* argv[])
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>( ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param); conv_param);
return run_conv_bwd_weight<3, return run_conv_bwd_data<3,
InDataType, InDataType,
WeiDataType, WeiDataType,
OutDataType, OutDataType,
InElementOp, InElementOp,
WeiElementOp, WeiElementOp,
OutElementOp, OutElementOp,
DeviceConvndBwdWeightInstance<3>>(do_verification, DeviceConvNdBwdDataInstance<3>>(do_verification,
init_method, init_method,
time_kernel, time_kernel,
conv_param, conv_param,
...@@ -208,8 +173,7 @@ int main(int argc, char* argv[]) ...@@ -208,8 +173,7 @@ int main(int argc, char* argv[])
out_g_n_k_wos_desc, out_g_n_k_wos_desc,
in_element_op, in_element_op,
wei_element_op, wei_element_op,
out_element_op, out_element_op);
split_k);
} }
return 0; return 0;
......
example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
View file @ f0224f2a
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
...@@ -132,15 +133,15 @@ int main(int argc, char* argv[]) ...@@ -132,15 +133,15 @@ int main(int argc, char* argv[])
std::size_t col, std::size_t col,
std::size_t stride, std::size_t stride,
auto layout) { auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value) if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}), return HostTensorDescriptor({batch_count, row, col}, {row * stride, stride, 1_uz});
std::vector<std::size_t>({row * stride, stride, 1}));
} }
else else
{ {
return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}), return HostTensorDescriptor({batch_count, row, col}, {col * stride, 1_uz, stride});
std::vector<std::size_t>({col * stride, 1, stride}));
} }
}; };
...@@ -149,17 +150,13 @@ int main(int argc, char* argv[]) ...@@ -149,17 +150,13 @@ int main(int argc, char* argv[])
Tensor<CDataType> c_g_m_n_host_result( Tensor<CDataType> c_g_m_n_host_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{})); f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>( Tensor<ReduceDataType> d0_g_m_host_result({BatchCount, M});
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)}))); Tensor<ReduceDataType> d1_g_m_host_result({BatchCount, M});
Tensor<ReduceDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
Tensor<CDataType> c_g_m_n_device_result( Tensor<CDataType> c_g_m_n_device_result(
f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{})); f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
Tensor<ReduceDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>( Tensor<ReduceDataType> d0_g_m_device_result({BatchCount, M});
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)}))); Tensor<ReduceDataType> d1_g_m_device_result({BatchCount, M});
Tensor<ReduceDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
{static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl; std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl; std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
...@@ -296,16 +293,15 @@ int main(int argc, char* argv[]) ...@@ -296,16 +293,15 @@ int main(int argc, char* argv[])
} }
} }
pass = ck::utils::check_err(c_g_m_n_host_result.mData, pass = ck::utils::check_err(
c_g_m_n_device_result.mData, c_g_m_n_host_result, c_g_m_n_device_result, "Error: Incorrect results c") &&
"Error: Incorrect results c") && ck::utils::check_err(d0_g_m_device_result,
ck::utils::check_err(d0_g_m_device_result.mData, d0_g_m_host_result,
d0_g_m_host_result.mData,
"Error: Incorrect results! D0", "Error: Incorrect results! D0",
1e-4, 1e-4,
1e-5) && 1e-5) &&
ck::utils::check_err(d1_g_m_device_result.mData, ck::utils::check_err(d1_g_m_device_result,
d1_g_m_host_result.mData, d1_g_m_host_result,
"Error: Incorrect results! D1", "Error: Incorrect results! D1",
1e-3, 1e-3,
1e-5); 1e-5);
......
example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
View file @ f0224f2a
...@@ -12,6 +12,7 @@ ...@@ -12,6 +12,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
using F16 = ck::half_t; using F16 = ck::half_t;
using F32 = float; using F32 = float;
...@@ -71,13 +72,13 @@ int main() ...@@ -71,13 +72,13 @@ int main()
ck::index_t Stride = 1024; ck::index_t Stride = 1024;
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) { auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}), return HostTensorDescriptor({len}, {stride});
std::vector<std::size_t>({stride}));
}; };
auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) { auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), using namespace ck::literals;
std::vector<std::size_t>({stride, 1}));
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}; };
Tensor<ABDataType> a_m_n(f_host_tensor_descriptor2d(M, N, Stride)); Tensor<ABDataType> a_m_n(f_host_tensor_descriptor2d(M, N, Stride));
...@@ -128,8 +129,7 @@ int main() ...@@ -128,8 +129,7 @@ int main()
host_broadcast2D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add, 0>( host_broadcast2D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add, 0>(
host_c_m_n, a_m_n, b_n, M, N, Add{}); host_c_m_n, a_m_n, b_n, M, N, Add{});
pass &= ck::utils::check_err( pass &= ck::utils::check_err(c_m_n, host_c_m_n, "Error: Incorrect results c", 1e-3, 1e-3);
c_m_n.mData, host_c_m_n.mData, "Error: Incorrect results c", 1e-3, 1e-3);
} }
return pass ? 0 : 1; return pass ? 0 : 1;
......
example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
View file @ f0224f2a
...@@ -8,6 +8,7 @@ ...@@ -8,6 +8,7 @@
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp" #include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
...@@ -82,11 +83,9 @@ int main() ...@@ -82,11 +83,9 @@ int main()
std::array<ck::index_t, 3> b_strides; std::array<ck::index_t, 3> b_strides;
std::array<ck::index_t, 3> c_strides; std::array<ck::index_t, 3> c_strides;
std::copy(mnk.begin(), mnk.end(), abc_lengths.begin()); ck::ranges::copy(mnk, abc_lengths.begin());
std::copy( ck::ranges::copy(b_m_n_k.mDesc.GetStrides(), b_strides.begin());
b_m_n_k.mDesc.GetStrides().begin(), b_m_n_k.mDesc.GetStrides().end(), b_strides.begin()); ck::ranges::copy(c_m_n_k.mDesc.GetStrides(), c_strides.begin());
std::copy(
c_m_n_k.mDesc.GetStrides().begin(), c_m_n_k.mDesc.GetStrides().end(), c_strides.begin());
auto broadcastAdd = DeviceElementwiseAddInstance{}; auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer( auto argument = broadcastAdd.MakeArgumentPointer(
...@@ -113,8 +112,8 @@ int main() ...@@ -113,8 +112,8 @@ int main()
host_broadcast3D_am_bmnk<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>( host_broadcast3D_am_bmnk<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
host_c_m_n_k, a_m, b_m_n_k, mnk, Add{}); host_c_m_n_k, a_m, b_m_n_k, mnk, Add{});
pass &= ck::utils::check_err( pass &=
c_m_n_k.mData, host_c_m_n_k.mData, "Error: Incorrect results c", 1e-3, 1e-3); ck::utils::check_err(c_m_n_k, host_c_m_n_k, "Error: Incorrect results c", 1e-3, 1e-3);
} }
return pass ? 0 : 1; return pass ? 0 : 1;
......
example/19_binary_elementwise/elementwise_add_1d.cpp
View file @ f0224f2a
...@@ -53,8 +53,7 @@ int main() ...@@ -53,8 +53,7 @@ int main()
ck::index_t M = 1024; ck::index_t M = 1024;
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) { auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}), return HostTensorDescriptor({len}, {stride});
std::vector<std::size_t>({stride}));
}; };
Tensor<ABDataType> a_m(f_host_tensor_descriptor1d(M, 1)); Tensor<ABDataType> a_m(f_host_tensor_descriptor1d(M, 1));
...@@ -105,8 +104,7 @@ int main() ...@@ -105,8 +104,7 @@ int main()
host_elementwise1D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>( host_elementwise1D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
host_c_m, a_m, b_m, M, Add{}); host_c_m, a_m, b_m, M, Add{});
pass &= ck::utils::check_err( pass &= ck::utils::check_err(c_m, host_c_m, "Error: Incorrect results c", 1e-3, 1e-3);
c_m.mData, host_c_m.mData, "Error: Incorrect results c", 1e-3, 1e-3);
} }
return pass ? 0 : 1; return pass ? 0 : 1;
......
example/19_binary_elementwise/elementwise_add_4d.cpp
View file @ f0224f2a
...@@ -8,6 +8,7 @@ ...@@ -8,6 +8,7 @@
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp" #include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
...@@ -82,10 +83,10 @@ int main() ...@@ -82,10 +83,10 @@ int main()
std::array<ck::index_t, 4> b_strides; std::array<ck::index_t, 4> b_strides;
std::array<ck::index_t, 4> c_strides; std::array<ck::index_t, 4> c_strides;
std::copy(nchw.begin(), nchw.end(), abc_lengths.begin()); ck::ranges::copy(nchw, abc_lengths.begin());
std::copy(a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end(), a_strides.begin()); ck::ranges::copy(a.mDesc.GetStrides(), a_strides.begin());
std::copy(b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end(), b_strides.begin()); ck::ranges::copy(b.mDesc.GetStrides(), b_strides.begin());
std::copy(c.mDesc.GetStrides().begin(), c.mDesc.GetStrides().end(), c_strides.begin()); ck::ranges::copy(c.mDesc.GetStrides(), c_strides.begin());
auto broadcastAdd = DeviceElementwiseAddInstance{}; auto broadcastAdd = DeviceElementwiseAddInstance{};
auto argument = broadcastAdd.MakeArgumentPointer( auto argument = broadcastAdd.MakeArgumentPointer(
...@@ -112,8 +113,7 @@ int main() ...@@ -112,8 +113,7 @@ int main()
host_elementwise4D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>( host_elementwise4D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
host_c, a, b, nchw, Add{}); host_c, a, b, nchw, Add{});
pass &= pass &= ck::utils::check_err(c, host_c, "Error: Incorrect results c", 1e-3, 1e-3);
ck::utils::check_err(c.mData, host_c.mData, "Error: Incorrect results c", 1e-3, 1e-3);
} }
return pass ? 0 : 1; return pass ? 0 : 1;
......
example/20_convnd_bwd_weight/CMakeLists.txt deleted 100644 → 0
View file @ befc2638
add_example_executable(example_convnd_bwd_weight_xdl_fp16 convnd_bwd_weight_xdl_fp16.cpp)
add_example_executable(example_convnd_bwd_weight_xdl_bf16 convnd_bwd_weight_xdl_bf16.cpp)
target_link_libraries(example_convnd_bwd_weight_xdl_fp16 PRIVATE utility)
target_link_libraries(example_convnd_bwd_weight_xdl_bf16 PRIVATE utility)
example/20_convnd_bwd_weight/convnd_bwd_weight_xdl_bf16.cpp deleted 100644 → 0
View file @ befc2638
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_bwd_weight_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_weight_nwc_kxc_nwk_xdl_cshuffle.hpp"
using InDataType = ck::bhalf_t;
// bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
using WeiDataType = float;
using OutDataType = ck::bhalf_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdWeightDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
template <ck::index_t NDimSpatial>
using DeviceConvndBwdWeightInstance =
ck::tensor_operation::device::DeviceConvNdBwdWeightNwcKxcNwk_Xdl_CShuffle<
NDimSpatial, // NDimSpatial
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
4>; // CBlockTransferScalarPerVector_NWaveNPerXdl
int main(int argc, char* argv[])
{
namespace ctc = ck::tensor_layout::convolution;
print_helper_msg();
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
ck::utils::conv::ConvParam conv_param{
2, 1, 32, 256, 1024, {3, 3}, {14, 14}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
ck::index_t split_k = 4;
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
split_k = std::stoi(argv[5 + 3 + 6 * num_dim_spatial - 1]);
split_k = std::max(1, split_k);
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(conv_param.num_dim_spatial_ == 1)
{
using InLayout = ctc::GNWC;
using WeiLayout = ctc::GKXC;
using OutLayout = ctc::GNWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_conv_bwd_weight<1,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<1>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
else if(conv_param.num_dim_spatial_ == 2)
{
using InLayout = ctc::GNHWC;
using WeiLayout = ctc::GKYXC;
using OutLayout = ctc::GNHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_conv_bwd_weight<2,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<2>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
else if(conv_param.num_dim_spatial_ == 3)
{
using InLayout = ctc::GNDHWC;
using WeiLayout = ctc::GKZYXC;
using OutLayout = ctc::GNDHWK;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_conv_bwd_weight<3,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceConvndBwdWeightInstance<3>>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op,
split_k);
}
return 0;
}
example/20_grouped_conv_bwd_weight/CMakeLists.txt 0 → 100644
View file @ f0224f2a
add_custom_target(example_grouped_conv_bwd_weight)
add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16
example_grouped_conv_bwd_weight_xdl_bf16)
example/20_grouped_conv_bwd_weight/common.hpp 0 → 100644
View file @ f0224f2a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <algorithm>
#include <iostream>
#include <iterator>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp"
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdWeightDefault =
ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
template <typename InputLay, typename WeightLay, typename OutputLay>
struct CommonLayoutSetting
{
using InputLayout = InputLay;
using WeightLayout = WeightLay;
using OutputLayout = OutputLay;
};
template <ck::index_t NDimSpatial>
struct CommonLayoutSettingSelector;
namespace ctl = ck::tensor_layout::convolution;
template <>
struct CommonLayoutSettingSelector<1> final : CommonLayoutSetting<ctl::GNWC, ctl::GKXC, ctl::GNWK>
{
};
template <>
struct CommonLayoutSettingSelector<2> final
: CommonLayoutSetting<ctl::GNHWC, ctl::GKYXC, ctl::GNHWK>
{
};
template <>
struct CommonLayoutSettingSelector<3> final
: CommonLayoutSetting<ctl::GNDHWC, ctl::GKZYXC, ctl::GNDHWK>
{
};
template <ck::index_t NDimSpatial>
using InputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::InputLayout;
template <ck::index_t NDimSpatial>
using WeightLayout = typename CommonLayoutSettingSelector<NDimSpatial>::WeightLayout;
template <ck::index_t NDimSpatial>
using OutputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::OutputLayout;
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
#define DefaultConvParam \
ck::utils::conv::ConvParam \
{ \
2, 4, 1, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, { 1, 1 } \
}
inline void print_help_msg()
{
std::cerr << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
}
inline bool parse_cmd_args(int argc,
char* argv[],
ExecutionConfig& config,
ck::utils::conv::ConvParam& conv_param)
{
constexpr int num_execution_config_args =
3; // arguments for do_verification, init_method, time_kernel
constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
constexpr int threshold_to_catch_all_args =
threshold_to_catch_partial_args + num_conv_param_leading_args;
if(argc == 1)
{
// use default
}
// catch only ExecutionConfig arguments
else if(argc == threshold_to_catch_partial_args)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
// catch both ExecutionConfig & ConvParam arguments
else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(
num_dim_spatial, threshold_to_catch_partial_args, argv);
}
else
{
print_help_msg();
return false;
}
return true;
}
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp 0 → 100644
View file @ f0224f2a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using InDataType = BF16;
// bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
using WeiDataType = F32;
using OutDataType = BF16;
using AccDataType = F32;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = PassThrough;
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp 0 → 100644
View file @ f0224f2a
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using InDataType = F16;
using WeiDataType = F16;
using OutDataType = F16;
using AccDataType = F32;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = PassThrough;
#include "run_grouped_conv_bwd_weight_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
example/20_convnd_bwd_weight/convnd_bwd_weight_common.hpp → example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
View file @ f0224f2a
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> template <ck::index_t NDimSpatial>
#include <numeric> using DeviceConvBwdWeightInstance =
#include <initializer_list> ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Xdl_CShuffle<
#include <cstdlib> NDimSpatial, // NDimSpatial
InDataType, // InDataType
#include "ck/ck.hpp" WeiDataType, // WeiDataType
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" OutDataType, // OutDataType
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
#include "ck/library/utility/check_err.hpp" WeiElementOp, // WeiElementwiseOperation
#include "ck/library/utility/device_memory.hpp" OutElementOp, // OutElementwiseOperation
#include "ck/library/utility/host_tensor.hpp" ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
#include "ck/library/utility/host_tensor_generator.hpp" 256, // BlockSize
#include "ck/library/utility/convolution_parameter.hpp" 128, // MPerBlock
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp" 128, // NPerBlock
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp" 4, // K0PerBlock
8, // K1
void print_helper_msg() 32, // MPerXdl
{ 32, // NPerXdl
std::cout << "arg1: verification (0=no, 1=yes)\n" 2, // MXdlPerWave
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n" 2, // NXdlPerWave
<< "arg3: time kernel (0=no, 1=yes)\n" S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl; S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
} S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
template <ck::index_t NDimSpatial>
using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
template <ck::index_t NDimSpatial, template <ck::index_t NDimSpatial>
typename InDataType, bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
typename WeiDataType, const ck::utils::conv::ConvParam& conv_param)
typename OutDataType,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp,
typename DeviceConvBwdWeightInstance>
int run_conv_bwd_weight(bool do_verification,
int init_method,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param,
const HostTensorDescriptor& in_g_n_c_wis_desc,
const HostTensorDescriptor& wei_g_k_c_xs_desc,
const HostTensorDescriptor& out_g_n_k_wos_desc,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op,
ck::index_t split_k)
{ {
constexpr ck::index_t split_k = 2;
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<
InputLayout<NDimSpatial>>(conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<
WeightLayout<NDimSpatial>>(conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<
OutputLayout<NDimSpatial>>(conv_param);
Tensor<InDataType> in(in_g_n_c_wis_desc); Tensor<InDataType> in(in_g_n_c_wis_desc);
Tensor<WeiDataType> wei_host_result(wei_g_k_c_xs_desc); Tensor<WeiDataType> wei_host_result(wei_g_k_c_xs_desc);
Tensor<WeiDataType> wei_device_result(wei_g_k_c_xs_desc); Tensor<WeiDataType> wei_device_result(wei_g_k_c_xs_desc);
...@@ -55,7 +77,7 @@ int run_conv_bwd_weight(bool do_verification, ...@@ -55,7 +77,7 @@ int run_conv_bwd_weight(bool do_verification,
std::cout << "wei: " << wei_host_result.mDesc << std::endl; std::cout << "wei: " << wei_host_result.mDesc << std::endl;
std::cout << "out: " << out.mDesc << std::endl; std::cout << "out: " << out.mDesc << std::endl;
switch(init_method) switch(config.init_method)
{ {
case 0: break; case 0: break;
case 1: case 1:
...@@ -77,36 +99,55 @@ int run_conv_bwd_weight(bool do_verification, ...@@ -77,36 +99,55 @@ int run_conv_bwd_weight(bool do_verification,
// init to 0 // init to 0
wei_device_buf.SetZero(); wei_device_buf.SetZero();
std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto range_copy = [](const auto& from, auto to) { std::copy(begin(from), end(from), to); };
range_copy(conv_param.input_spatial_lengths_, begin(input_spatial_lengths));
range_copy(conv_param.filter_spatial_lengths_, begin(filter_spatial_lengths));
range_copy(conv_param.output_spatial_lengths_, begin(output_spatial_lengths));
range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
range_copy(conv_param.input_left_pads_, begin(input_left_pads));
range_copy(conv_param.input_right_pads_, begin(input_right_pads));
// do GEMM // do GEMM
auto conv = DeviceConvBwdWeightInstance{}; auto conv = DeviceConvBwdWeightInstance<NDimSpatial>{};
auto invoker = conv.MakeInvoker(); auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()), auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()), static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()), static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
conv_param.G_,
conv_param.N_, conv_param.N_,
conv_param.K_, conv_param.K_,
conv_param.C_, conv_param.C_,
conv_param.input_spatial_lengths_, input_spatial_lengths,
conv_param.filter_spatial_lengths_, filter_spatial_lengths,
conv_param.output_spatial_lengths_, output_spatial_lengths,
conv_param.conv_filter_strides_, conv_filter_strides,
conv_param.conv_filter_dilations_, conv_filter_dilations,
conv_param.input_left_pads_, input_left_pads,
conv_param.input_right_pads_, input_right_pads,
in_element_op, InElementOp{},
wei_element_op, WeiElementOp{},
out_element_op, OutElementOp{},
split_k); split_k);
if(!conv.IsSupportedArgument(argument)) if(!conv.IsSupportedArgument(argument))
{ {
std::cout << "wrong! device_conv with the specified compilation parameters does " std::cerr << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem" "not support this Conv problem"
<< std::endl; << std::endl;
return 1; return false;
} }
float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
std::size_t flop = conv_param.GetFlops(); std::size_t flop = conv_param.GetFlops();
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>(); std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
...@@ -115,21 +156,14 @@ int run_conv_bwd_weight(bool do_verification, ...@@ -115,21 +156,14 @@ int run_conv_bwd_weight(bool do_verification,
float gb_per_sec = num_btype / 1.E6 / 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" std::cerr << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< conv.GetTypeString() << std::endl; << std::endl
<< "DeviceOp: " << conv.GetTypeString() << std::endl;
if(do_verification) if(config.do_verification)
{ {
auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial, auto ref_conv = HostConvBwdWeightInstance<NDimSpatial>{};
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>{};
auto ref_invoker = ref_conv.MakeInvoker(); auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in, auto ref_argument = ref_conv.MakeArgument(in,
wei_host_result, wei_host_result,
out, out,
...@@ -145,8 +179,28 @@ int run_conv_bwd_weight(bool do_verification, ...@@ -145,8 +179,28 @@ int run_conv_bwd_weight(bool do_verification,
wei_device_buf.FromDevice(wei_device_result.mData.data()); wei_device_buf.FromDevice(wei_device_result.mData.data());
return ck::utils::check_err(wei_device_result.mData, wei_host_result.mData) ? 0 : 1; return ck::utils::check_err(wei_device_result.mData, wei_host_result.mData);
}
return true;
}
bool run_grouped_conv_bwd_weight_example(int argc, char* argv[])
{
ExecutionConfig config;
ck::utils::conv::ConvParam conv_param = DefaultConvParam;
if(!parse_cmd_args(argc, argv, config, conv_param))
{
return false;
}
switch(conv_param.num_dim_spatial_)
{
case 1: return run_grouped_conv_bwd_weight<1>(config, conv_param);
case 2: return run_grouped_conv_bwd_weight<2>(config, conv_param);
case 3: return run_grouped_conv_bwd_weight<3>(config, conv_param);
} }
return 0; return false;
} }
example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_fp16.cpp
View file @ f0224f2a
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -108,21 +109,20 @@ using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise< ...@@ -108,21 +109,20 @@ using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
ck::Sequence<8>>; // scalarPerVector: y(layerNorm_out) ck::Sequence<8>>; // scalarPerVector: y(layerNorm_out)
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) { auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}), return HostTensorDescriptor({len}, {stride});
std::vector<std::size_t>({stride}));
}; };
auto f_host_tensor_descriptor2d = auto f_host_tensor_descriptor2d =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) { [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value) if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {stride, 1_uz});
std::vector<std::size_t>({stride, 1}));
} }
else else
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {1_uz, stride});
std::vector<std::size_t>({1, stride}));
} }
}; };
...@@ -372,8 +372,8 @@ int main() ...@@ -372,8 +372,8 @@ int main()
N); N);
layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data()); layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
pass &= ck::utils::check_err(layerNorm_m_n.mData, pass &= ck::utils::check_err(layerNorm_m_n,
host_layerNorm_m_n.mData, host_layerNorm_m_n,
"Error: Incorrect results layerNorm_m_n", "Error: Incorrect results layerNorm_m_n",
1e-2, 1e-2,
1e-2); 1e-2);
......
example/21_gemm_layernorm/gemm_layernorm_xdl_fp16.cpp
View file @ f0224f2a
...@@ -16,6 +16,7 @@ ...@@ -16,6 +16,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -107,21 +108,20 @@ using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise< ...@@ -107,21 +108,20 @@ using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwise<
ck::Sequence<8>>; // scalarPerVector: y(layerNorm_out) ck::Sequence<8>>; // scalarPerVector: y(layerNorm_out)
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) { auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}), return HostTensorDescriptor({len}, {stride});
std::vector<std::size_t>({stride}));
}; };
auto f_host_tensor_descriptor2d = auto f_host_tensor_descriptor2d =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) { [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value) if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {stride, 1_uz});
std::vector<std::size_t>({stride, 1}));
} }
else else
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {1_uz, stride});
std::vector<std::size_t>({1, stride}));
} }
}; };
...@@ -346,11 +346,8 @@ int main() ...@@ -346,11 +346,8 @@ int main()
N); N);
layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data()); layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
pass &= ck::utils::check_err(layerNorm_m_n.mData, pass &= ck::utils::check_err(
host_layerNorm_m_n.mData, layerNorm_m_n, host_layerNorm_m_n, "Error: Incorrect results d1", 1e-3, 1e-3);
"Error: Incorrect results d1",
1e-3,
1e-3);
} }
{ {
......
example/21_gemm_layernorm/gemm_xdl_layernorm_single_kernel_fp16.cpp
View file @ f0224f2a
...@@ -10,6 +10,7 @@ ...@@ -10,6 +10,7 @@
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
...@@ -132,15 +133,15 @@ int main(int argc, char* argv[]) ...@@ -132,15 +133,15 @@ int main(int argc, char* argv[])
auto f_host_tensor_descriptor = auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) { [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value) if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {stride, 1_uz});
std::vector<std::size_t>({stride, 1}));
} }
else else
{ {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), return HostTensorDescriptor({row, col}, {1_uz, stride});
std::vector<std::size_t>({1, stride}));
} }
}; };
...@@ -149,10 +150,10 @@ int main(int argc, char* argv[]) ...@@ -149,10 +150,10 @@ int main(int argc, char* argv[])
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<AccDataType> acc_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<AccDataType> acc_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<C0DataType> c0_n_bias(HostTensorDescriptor(std::vector<size_t>({size_t(N)}))); Tensor<C0DataType> c0_n_bias({N});
Tensor<C0DataType> c0_m_n_add(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<C0DataType> c0_m_n_add(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<C0DataType> c0_n_gamma(HostTensorDescriptor(std::vector<size_t>({size_t(N)}))); Tensor<C0DataType> c0_n_gamma({N});
Tensor<C0DataType> c0_n_beta(HostTensorDescriptor(std::vector<size_t>({size_t(N)}))); Tensor<C0DataType> c0_n_beta({N});
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl; std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
...@@ -274,15 +275,12 @@ int main(int argc, char* argv[]) ...@@ -274,15 +275,12 @@ int main(int argc, char* argv[])
if constexpr(std::is_same<CShuffleDataType, F32>::value) if constexpr(std::is_same<CShuffleDataType, F32>::value)
{ {
pass &= ck::utils::check_err( pass &= ck::utils::check_err(
c_m_n_device_result.mData, c_m_n_host_result.mData, "Error: Incorrect results c"); c_m_n_device_result, c_m_n_host_result, "Error: Incorrect results c");
} }
else if constexpr(std::is_same<CShuffleDataType, F16>::value) else if constexpr(std::is_same<CShuffleDataType, F16>::value)
{ {
pass &= ck::utils::check_err(c_m_n_device_result.mData, pass &= ck::utils::check_err(
c_m_n_host_result.mData, c_m_n_device_result, c_m_n_host_result, "Error: Incorrect results c", 1e-2, 1e-2);
"Error: Incorrect results c",
1e-2,
1e-2);
} }
} }
return pass ? 0 : 1; return pass ? 0 : 1;
......
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