Skip to content

GitLab

Commit 578ffb6b authored by Chao Liu's avatar Chao Liu
Browse files

update example

parent 5816a647
Show whitespace changes
Inline Side-by-side
Showing with 213 additions and 168 deletions
example/01_gemm/gemm_xdl_fp16.cpp
View file @ 578ffb6b
...@@ -27,28 +27,29 @@ using Col = ck::tensor_layout::gemm::ColumnMajor; ...@@ -27,28 +27,29 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = ck::half_t; using ADataType = F16;
using BDataType = ck::half_t; using BDataType = F16;
using CDataType = ck::half_t; using AccDataType = F32;
using AccDataType = float; using CShuffleDataType = F32;
using CDataType = F16;
using ALayout = ck::tensor_layout::gemm::RowMajor; using ALayout = Row;
using BLayout = ck::tensor_layout::gemm::ColumnMajor; using BLayout = Col;
using CLayout = ck::tensor_layout::gemm::RowMajor; using CLayout = Row;
using AElementOp = ck::tensor_operation::element_wise::PassThrough; using AElementOp = PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough; using BElementOp = PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off // clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
//######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| //######| ALayout| BLayout| ELayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| //######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| //######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>; < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host:: using ReferenceGemmInstance = ck::tensor_operation::host::
...@@ -69,7 +70,11 @@ int main(int argc, char* argv[]) ...@@ -69,7 +70,11 @@ int main(int argc, char* argv[])
ck::index_t StrideB = 4096; ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096; ck::index_t StrideC = 4096;
if(argc == 4) if(argc == 1)
{
// use default case
}
else if(argc == 4)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
...@@ -93,7 +98,7 @@ int main(int argc, char* argv[]) ...@@ -93,7 +98,7 @@ int main(int argc, char* argv[])
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n"); printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"); printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0); exit(0);
} }
......
example/03_gemm_bias_add_fastgelu/gemm_bias_add_fastgelu_xdl_fp16.cpp
View file @ 578ffb6b
...@@ -10,10 +10,10 @@ ...@@ -10,10 +10,10 @@
#include "host_tensor.hpp" #include "host_tensor.hpp"
#include "host_tensor_generator.hpp" #include "host_tensor_generator.hpp"
#include "device_tensor.hpp" #include "device_tensor.hpp"
#include "device_gemm_multiple_d_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp" #include "element_wise_operation.hpp"
#include "reference_gemm.hpp" #include "reference_gemm.hpp"
#include "gemm_specialization.hpp" #include "gemm_specialization.hpp"
#include "device_gemm_multiple_d_xdl_cshuffle.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
...@@ -26,8 +26,8 @@ using Col = ck::tensor_layout::gemm::ColumnMajor; ...@@ -26,8 +26,8 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// E = FastGelu((A * B) + D0 + D1)
// C = A * B // C = A * B
// E = FastGelu(C + D0 + D1)
struct AddAddFastGelu struct AddAddFastGelu
{ {
__host__ __device__ void __host__ __device__ void
...@@ -69,7 +69,7 @@ using CDEElementOp = AddAddFastGelu; ...@@ -69,7 +69,7 @@ using CDEElementOp = AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off // clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
//######| ALayout| BLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| //######| ALayout| BLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| //######| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| //######| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
...@@ -90,6 +90,7 @@ int main(int argc, char* argv[]) ...@@ -90,6 +90,7 @@ int main(int argc, char* argv[])
ck::index_t StrideA = 4096; ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096; ck::index_t StrideB = 4096;
ck::index_t StrideD1 = 4096;
ck::index_t StrideE = 4096; ck::index_t StrideE = 4096;
if(argc == 1) if(argc == 1)
...@@ -102,7 +103,7 @@ int main(int argc, char* argv[]) ...@@ -102,7 +103,7 @@ int main(int argc, char* argv[])
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]); time_kernel = std::stoi(argv[3]);
} }
else if(argc == 10) else if(argc == 11)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
...@@ -114,14 +115,15 @@ int main(int argc, char* argv[]) ...@@ -114,14 +115,15 @@ int main(int argc, char* argv[])
StrideA = std::stoi(argv[7]); StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]); StrideB = std::stoi(argv[8]);
StrideE = std::stoi(argv[9]); StrideD1 = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
} }
else else
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n"); printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"); printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD1, StrideE\n");
exit(0); exit(0);
} }
...@@ -184,30 +186,28 @@ int main(int argc, char* argv[]) ...@@ -184,30 +186,28 @@ int main(int argc, char* argv[])
auto cde_element_op = CDEElementOp{}; auto cde_element_op = CDEElementOp{};
// do GEMM // do GEMM
auto gemm = DeviceGemmInstance{}; auto device_op = DeviceOpInstance{};
auto invoker = gemm.MakeInvoker(); auto invoker = device_op.MakeInvoker();
auto argument = auto argument =
gemm.MakeArgument(a_m_k_device_buf.GetDeviceBuffer(), device_op.MakeArgument(a_m_k_device_buf.GetDeviceBuffer(),
b_k_n_device_buf.GetDeviceBuffer(), b_k_n_device_buf.GetDeviceBuffer(),
std::array<const void*, 2>{{d0_m_n_device_buf.GetDeviceBuffer(), std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
d1_m_n_device_buf.GetDeviceBuffer()}}, d1_m_n_device_buf.GetDeviceBuffer()},
e_m_n_device_buf.GetDeviceBuffer(), e_m_n_device_buf.GetDeviceBuffer(),
M, M,
N, N,
K, K,
StrideA, StrideA,
StrideB, StrideB,
std::array<ck::index_t, 2>{{0, StrideE}}, std::array<ck::index_t, 2>{0, StrideD1},
StrideE, StrideE,
a_element_op, a_element_op,
b_element_op, b_element_op,
cde_element_op); cde_element_op);
if(!gemm.IsSupportedArgument(argument)) if(!device_op.IsSupportedArgument(argument))
{ {
throw std::runtime_error( throw std::runtime_error("wrong! this device_op instance does not support this problem");
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
...@@ -222,7 +222,7 @@ int main(int argc, char* argv[]) ...@@ -222,7 +222,7 @@ int main(int argc, char* argv[])
float gb_per_sec = num_btype / 1.E6 / 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, " std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl; << device_op.GetTypeString() << std::endl;
if(do_verification) if(do_verification)
{ {
...@@ -237,6 +237,7 @@ int main(int argc, char* argv[]) ...@@ -237,6 +237,7 @@ int main(int argc, char* argv[])
AElementOp, AElementOp,
BElementOp, BElementOp,
PassThrough>; PassThrough>;
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
......
example/03_gemm_bias_relu/gemm_xdl_bias_relu.cpp
View file @ 578ffb6b
...@@ -3,83 +3,103 @@ ...@@ -3,83 +3,103 @@
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h> #include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp" #include "check_err.hpp"
#include "config.hpp" #include "config.hpp"
#include "print.hpp"
#include "device.hpp" #include "device.hpp"
#include "host_tensor.hpp" #include "host_tensor.hpp"
#include "host_tensor_generator.hpp" #include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp" #include "device_tensor.hpp"
#include "element_wise_operation.hpp" #include "element_wise_operation.hpp"
#include "device_gemm_xdl_c_shuffle_bias_activation.hpp" #include "reference_gemm.hpp"
#include "reference_gemm_bias_activation.hpp" #include "gemm_specialization.hpp"
#include "device_gemm_multiple_d_xdl_cshuffle.hpp"
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
using ADataType = ck::half_t; using F16 = ck::half_t;
using BDataType = ck::half_t; using F32 = float;
using CDataType = ck::half_t;
using AccDataType = float; using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using CLayout = ck::tensor_layout::gemm::RowMajor;
// C = A * B
using AElementOp = ck::tensor_operation::element_wise::PassThrough; // E = Relu(C + D);
using BElementOp = ck::tensor_operation::element_wise::PassThrough; struct AddRelu
using CElementOp = ck::tensor_operation::element_wise::AddRelu; {
__host__ __device__ void
// clang-format off operator()(ck::half_t& e, const ck::half_t& c, const ck::half_t& d) const
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle_Bias_Activation< {
ADataType, // ADataType const ck::half_t x = c + d;
BDataType, // BDataType
CDataType, // CDataType e = x > 0 ? x : 0;
AccDataType, // AccDataType }
ALayout, // ALayout };
BLayout, // BLayout
CLayout, // CLayout using ADataType = F16;
AElementOp, // AElementwiseOperation using BDataType = F16;
BElementOp, // BElementwiseOperation using AccDataType = F32;
CElementOp, // CElementwiseOperation using CShuffleDataType = F16;
256, // BlockSize using DDataType = F16;
256, // MPerBlock using DsDataType = ck::Tuple<DDataType>;
128, // NPerBlock using EDataType = F16;
4, // K0PerBlock
8, // K1 using ALayout = Row;
32, // MPerXDL using BLayout = Col;
32, // NPerXDL using ELayout = Row;
4, // MXdlPerWave
2, // NXdlPerWave using AElementOp = PassThrough;
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1 using BElementOp = PassThrough;
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder using CDEElementOp = AddRelu;
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1 using DeviceOpInstance =
true, // ABlockLdsAddExtraM ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<ALayout,
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1 BLayout,
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder ELayout,
S<1, 0, 2>, // BBlockTransferSrcAccessOrder ADataType,
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 1, 32, 1, 1, 8>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBiasActivation<ADataType,
BDataType, BDataType,
CDataType, AccDataType,
CShuffleDataType,
DsDataType,
EDataType,
AElementOp, AElementOp,
BElementOp, BElementOp,
CElementOp>; CDEElementOp,
GemmDefault,
1,
256,
256,
128,
32,
8,
8,
32,
32,
4,
2,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
1,
1,
S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
...@@ -94,9 +114,13 @@ int main(int argc, char* argv[]) ...@@ -94,9 +114,13 @@ int main(int argc, char* argv[])
ck::index_t StrideA = 4096; ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096; ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096; ck::index_t StrideE = 4096;
if(argc == 4) if(argc == 1)
{
// use default case
}
else if(argc == 4)
{ {
do_verification = std::stoi(argv[1]); do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); init_method = std::stoi(argv[2]);
...@@ -114,14 +138,14 @@ int main(int argc, char* argv[]) ...@@ -114,14 +138,14 @@ int main(int argc, char* argv[])
StrideA = std::stoi(argv[7]); StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]); StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]); StrideE = std::stoi(argv[9]);
} }
else else
{ {
printf("arg1: verification (0=no, 1=yes)\n"); printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n"); printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"); printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n");
exit(0); exit(0);
} }
...@@ -141,17 +165,14 @@ int main(int argc, char* argv[]) ...@@ -141,17 +165,14 @@ int main(int argc, char* argv[])
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{})); Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, 0, ELayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
// c0_n[n]
Tensor<CDataType> c0_n(HostTensorDescriptor(
std::vector<std::size_t>({static_cast<std::size_t>(N)}), std::vector<std::size_t>({1})));
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;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl; std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "c0_n: " << c0_n.mDesc << std::endl; std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method) switch(init_method)
{ {
...@@ -159,59 +180,59 @@ int main(int argc, char* argv[]) ...@@ -159,59 +180,59 @@ int main(int argc, char* argv[])
case 1: case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}); a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}); b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5}); d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
break; break;
default: default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}); a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}); b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0}); d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
} }
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace()); DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace()); DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace()); DeviceMem d_m_n_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpace());
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace()); DeviceMem e_m_n_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(a_m_k.mData.data()); a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data()); b_k_n_device_buf.ToDevice(b_k_n.mData.data());
c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data()); d_m_n_device_buf.ToDevice(d_m_n.mData.data());
c0_n_device_buf.ToDevice(c0_n.mData.data());
auto a_element_op = AElementOp{}; auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{}; auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{}; auto cde_element_op = CDEElementOp{};
// do GEMM // do GEMM
auto gemm = DeviceGemmInstance{}; auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto invoker = gemm.MakeInvoker(); auto argument =
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()), device_op.MakeArgument(a_m_k_device_buf.GetDeviceBuffer(),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()), b_k_n_device_buf.GetDeviceBuffer(),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()), std::array<const void*, 1>{d_m_n_device_buf.GetDeviceBuffer()},
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()), e_m_n_device_buf.GetDeviceBuffer(),
M, M,
N, N,
K, K,
StrideA, StrideA,
StrideB, StrideB,
StrideC, std::array<ck::index_t, 1>{0},
StrideE,
a_element_op, a_element_op,
b_element_op, b_element_op,
c_element_op); cde_element_op);
if(!gemm.IsSupportedArgument(argument)) if(!device_op.IsSupportedArgument(argument))
{ {
throw std::runtime_error( throw std::runtime_error("wrong! this device_op instance does not support this problem");
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
} }
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K; std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M + std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(CDataType) * N; sizeof(EDataType) * M * N + sizeof(EDataType) * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
...@@ -220,19 +241,37 @@ int main(int argc, char* argv[]) ...@@ -220,19 +241,37 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s" std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl; << std::endl;
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(do_verification) if(do_verification)
{ {
e_m_n_device_buf.FromDevice(e_m_n_device_result.mData.data());
Tensor<AccDataType> c_m_n(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument( auto ref_argument =
a_m_k, b_k_n, c_m_n_host_result, c0_n, a_element_op, b_element_op, c_element_op); ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1; for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
}
}
return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData) ? 0 : 1;
} }
return 0; return 0;
......
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