mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-15 18:42:06 +00:00
GEMM with Multiple Source, GEMM+Bias+Add+FastGeLU example and ckProfiler (#241)
* ad gelu and fast_gelu
* added GeLU and fast GeLU
* clean up
* add gemm+fastgelu example
* add gemm+gelu instances
* update profiler
* clean up
* clean up
* adding gemm+bias+activation
* clean
* adding bias
* clean
* adding gemm multiple d
* debugging
* add gemm bias add fastgelu
* rename, clean
* refactoring; add readme
* refactor
* refactor
* refactor
* refactor
* refactor
* refactor
* fix
* fix
* update example
* update example
* rename
* update example
* add ckProfiler
* clean
* clean
* clean
* clean
* add comment
* use type_convert
* clean
* clean element wise op
[ROCm/composable_kernel commit: 56adf7e9cc]
This commit is contained in:
Chao Liu
@@ -27,28 +27,29 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
using ADataType = ck::half_t;
|
||||
using BDataType = ck::half_t;
|
||||
using CDataType = ck::half_t;
|
||||
using AccDataType = float;
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using AccDataType = F32;
|
||||
using CShuffleDataType = F32;
|
||||
using CDataType = F16;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using CLayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using ALayout = Row;
|
||||
using BLayout = Col;
|
||||
using CLayout = Row;
|
||||
|
||||
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CElementOp = PassThrough;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
// clang-format off
|
||||
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|
|
||||
//######| | | | 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|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< 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| 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|
|
||||
//######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< ALayout, BLayout, 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
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::
|
||||
@@ -69,7 +70,11 @@ int main(int argc, char* argv[])
|
||||
ck::index_t StrideB = 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]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
@@ -93,7 +98,7 @@ int main(int argc, char* argv[])
|
||||
{
|
||||
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=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");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
@@ -3,83 +3,103 @@
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
#include <stdlib.h>
|
||||
#include <half.hpp>
|
||||
|
||||
#include "check_err.hpp"
|
||||
#include "config.hpp"
|
||||
#include "print.hpp"
|
||||
#include "device.hpp"
|
||||
#include "host_tensor.hpp"
|
||||
#include "host_tensor_generator.hpp"
|
||||
#include "host_gemm.hpp"
|
||||
#include "device_tensor.hpp"
|
||||
#include "element_wise_operation.hpp"
|
||||
#include "device_gemm_xdl_c_shuffle_bias_activation.hpp"
|
||||
#include "reference_gemm_bias_activation.hpp"
|
||||
#include "reference_gemm.hpp"
|
||||
#include "gemm_specialization.hpp"
|
||||
#include "device_gemm_multiple_d_xdl_cshuffle.hpp"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using ADataType = ck::half_t;
|
||||
using BDataType = ck::half_t;
|
||||
using CDataType = ck::half_t;
|
||||
using AccDataType = float;
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using CLayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using CElementOp = ck::tensor_operation::element_wise::AddRelu;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
|
||||
// clang-format off
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle_Bias_Activation<
|
||||
ADataType, // ADataType
|
||||
BDataType, // BDataType
|
||||
CDataType, // CDataType
|
||||
AccDataType, // AccDataType
|
||||
ALayout, // ALayout
|
||||
BLayout, // BLayout
|
||||
CLayout, // CLayout
|
||||
AElementOp, // AElementwiseOperation
|
||||
BElementOp, // BElementwiseOperation
|
||||
CElementOp, // CElementwiseOperation
|
||||
256, // BlockSize
|
||||
256, // MPerBlock
|
||||
128, // NPerBlock
|
||||
4, // K0PerBlock
|
||||
8, // K1
|
||||
32, // MPerXDL
|
||||
32, // NPerXDL
|
||||
4, // MXdlPerWave
|
||||
2, // NXdlPerWave
|
||||
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
|
||||
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
|
||||
2, // ABlockTransferSrcVectorDim
|
||||
8, // ABlockTransferSrcScalarPerVector
|
||||
8, // ABlockTransferDstScalarPerVector_K1
|
||||
true, // ABlockLdsAddExtraM
|
||||
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
|
||||
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
|
||||
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
|
||||
// C = A * B
|
||||
// E = Relu(C + D);
|
||||
struct AddRelu
|
||||
{
|
||||
__host__ __device__ void
|
||||
operator()(ck::half_t& e, const ck::half_t& c, const ck::half_t& d) const
|
||||
{
|
||||
const ck::half_t x = c + d;
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBiasActivation<ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CElementOp>;
|
||||
e = x > 0 ? x : 0;
|
||||
}
|
||||
};
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using AccDataType = F32;
|
||||
using CShuffleDataType = F16;
|
||||
using DDataType = F16;
|
||||
using DsDataType = ck::Tuple<DDataType>;
|
||||
using EDataType = F16;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Col;
|
||||
using ELayout = Row;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CDEElementOp = AddRelu;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
using DeviceOpInstance =
|
||||
ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<ALayout,
|
||||
BLayout,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
CDEElementOp,
|
||||
GemmDefault,
|
||||
1,
|
||||
256,
|
||||
256,
|
||||
128,
|
||||
32,
|
||||
8,
|
||||
8,
|
||||
32,
|
||||
32,
|
||||
4,
|
||||
2,
|
||||
S<4, 64, 1>,
|
||||
S<1, 0, 2>,
|
||||
S<1, 0, 2>,
|
||||
2,
|
||||
8,
|
||||
8,
|
||||
1,
|
||||
S<4, 64, 1>,
|
||||
S<1, 0, 2>,
|
||||
S<1, 0, 2>,
|
||||
2,
|
||||
8,
|
||||
8,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
S<1, 32, 1, 8>,
|
||||
8>;
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
@@ -94,9 +114,13 @@ int main(int argc, char* argv[])
|
||||
|
||||
ck::index_t StrideA = 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]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
@@ -114,14 +138,14 @@ int main(int argc, char* argv[])
|
||||
|
||||
StrideA = std::stoi(argv[7]);
|
||||
StrideB = std::stoi(argv[8]);
|
||||
StrideC = std::stoi(argv[9]);
|
||||
StrideE = std::stoi(argv[9]);
|
||||
}
|
||||
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=n0, 1=yes)\n");
|
||||
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
|
||||
printf("arg3: time kernel (0=no, 1=yes)\n");
|
||||
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
@@ -141,17 +165,14 @@ int main(int argc, char* argv[])
|
||||
|
||||
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<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{}));
|
||||
|
||||
// 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})));
|
||||
Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, 0, ELayout{}));
|
||||
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{}));
|
||||
|
||||
std::cout << "a_m_k: " << a_m_k.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 << "c0_n: " << c0_n.mDesc << std::endl;
|
||||
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
|
||||
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
@@ -159,59 +180,59 @@ int main(int argc, char* argv[])
|
||||
case 1:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-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;
|
||||
default:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
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 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 c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
|
||||
DeviceMem d_m_n_device_buf(sizeof(DDataType) * d_m_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());
|
||||
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
|
||||
c_m_n_device_buf.ToDevice(c_m_n_device_result.mData.data());
|
||||
c0_n_device_buf.ToDevice(c0_n.mData.data());
|
||||
d_m_n_device_buf.ToDevice(d_m_n.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto c_element_op = CElementOp{};
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto cde_element_op = CDEElementOp{};
|
||||
|
||||
// do GEMM
|
||||
auto gemm = DeviceGemmInstance{};
|
||||
auto device_op = DeviceOpInstance{};
|
||||
|
||||
auto invoker = gemm.MakeInvoker();
|
||||
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
|
||||
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op);
|
||||
auto invoker = device_op.MakeInvoker();
|
||||
|
||||
if(!gemm.IsSupportedArgument(argument))
|
||||
auto argument =
|
||||
device_op.MakeArgument(a_m_k_device_buf.GetDeviceBuffer(),
|
||||
b_k_n_device_buf.GetDeviceBuffer(),
|
||||
std::array<const void*, 1>{d_m_n_device_buf.GetDeviceBuffer()},
|
||||
e_m_n_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
std::array<ck::index_t, 1>{0},
|
||||
StrideE,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
if(!device_op.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
throw std::runtime_error("wrong! this device_op instance does not support this problem");
|
||||
}
|
||||
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
|
||||
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
|
||||
sizeof(CDataType) * M * N + sizeof(CDataType) * N;
|
||||
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
|
||||
sizeof(EDataType) * M * N + sizeof(EDataType) * N;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
|
||||
@@ -220,19 +241,37 @@ int main(int argc, char* argv[])
|
||||
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
|
||||
<< std::endl;
|
||||
|
||||
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
|
||||
|
||||
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_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(
|
||||
a_m_k, b_k_n, c_m_n_host_result, c0_n, a_element_op, b_element_op, c_element_op);
|
||||
auto ref_argument =
|
||||
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
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;
|
||||
|
||||
1
example/04_gemm_add_add_fastgelu/CMakeLists.txt
Normal file
1
example/04_gemm_add_add_fastgelu/CMakeLists.txt
Normal file
@@ -0,0 +1 @@
|
||||
add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
|
||||
23
example/04_gemm_add_add_fastgelu/README.md
Normal file
23
example/04_gemm_add_add_fastgelu/README.md
Normal file
@@ -0,0 +1,23 @@
|
||||
# Instructions for ```example_gemm_add_add_fastgelu_xdl_fp16```
|
||||
|
||||
## Run ```example_gemm_add_add_fastgelu_xdl_fp16```
|
||||
```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)
|
||||
#arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
|
||||
./bin/example_gemm_add_add_fastgelu_xdl_fp16 1 1 1
|
||||
```
|
||||
|
||||
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
|
||||
```
|
||||
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
|
||||
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
|
||||
d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
|
||||
d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
|
||||
e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
|
||||
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
|
||||
Warm up 1 time
|
||||
Start running 10 times...
|
||||
Perf: 1.26914 ms, 101.525 TFlops, 100.804 GB/s, DeviceGemmMultipleD_Xdl_CShuffle<256, 256, 128, 32, 8, 8>
|
||||
```
|
||||
@@ -0,0 +1,245 @@
|
||||
#include <iostream>
|
||||
#include <numeric>
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
#include <stdlib.h>
|
||||
|
||||
#include "check_err.hpp"
|
||||
#include "config.hpp"
|
||||
#include "device.hpp"
|
||||
#include "host_tensor.hpp"
|
||||
#include "host_tensor_generator.hpp"
|
||||
#include "device_tensor.hpp"
|
||||
#include "element_wise_operation.hpp"
|
||||
#include "reference_gemm.hpp"
|
||||
#include "gemm_specialization.hpp"
|
||||
#include "device_gemm_multiple_d_xdl_cshuffle.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 AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using AccDataType = F32;
|
||||
using CShuffleDataType = F32;
|
||||
using D0DataType = F16;
|
||||
using D1DataType = F16;
|
||||
using DsDataType = ck::Tuple<D0DataType, D1DataType>;
|
||||
using EDataType = F16;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Col;
|
||||
using D0Layout = Row;
|
||||
using D1Layout = Row;
|
||||
using ELayout = Row;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = PassThrough;
|
||||
using CDEElementOp = AddAddFastGelu;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
// clang-format off
|
||||
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|
|
||||
//######| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
|
||||
//######| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< ALayout, BLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>;
|
||||
// clang-format on
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 1;
|
||||
bool time_kernel = false;
|
||||
|
||||
// GEMM shape
|
||||
ck::index_t M = 3840;
|
||||
ck::index_t N = 4096;
|
||||
ck::index_t K = 4096;
|
||||
|
||||
ck::index_t StrideA = 4096;
|
||||
ck::index_t StrideB = 4096;
|
||||
ck::index_t StrideD0 = 0;
|
||||
ck::index_t StrideD1 = 4096;
|
||||
ck::index_t StrideE = 4096;
|
||||
|
||||
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 == 12)
|
||||
{
|
||||
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]);
|
||||
|
||||
StrideA = std::stoi(argv[7]);
|
||||
StrideB = std::stoi(argv[8]);
|
||||
StrideD0 = std::stoi(argv[9]);
|
||||
StrideD1 = std::stoi(argv[10]);
|
||||
StrideE = std::stoi(argv[11]);
|
||||
}
|
||||
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 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, "
|
||||
"StrideE\n");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
auto f_host_tensor_descriptor =
|
||||
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
||||
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
|
||||
std::vector<std::size_t>({stride, 1}));
|
||||
}
|
||||
else
|
||||
{
|
||||
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
|
||||
std::vector<std::size_t>({1, stride}));
|
||||
}
|
||||
};
|
||||
|
||||
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<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
|
||||
Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
|
||||
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{}));
|
||||
|
||||
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
|
||||
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
|
||||
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
|
||||
std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
|
||||
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
|
||||
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
|
||||
d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-5, 5});
|
||||
break;
|
||||
default:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
|
||||
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
|
||||
}
|
||||
|
||||
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
|
||||
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
|
||||
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpace());
|
||||
DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_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());
|
||||
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
|
||||
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
|
||||
d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto cde_element_op = CDEElementOp{};
|
||||
|
||||
// do GEMM
|
||||
auto device_op = DeviceOpInstance{};
|
||||
auto invoker = device_op.MakeInvoker();
|
||||
auto argument =
|
||||
device_op.MakeArgument(a_m_k_device_buf.GetDeviceBuffer(),
|
||||
b_k_n_device_buf.GetDeviceBuffer(),
|
||||
std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
|
||||
d1_m_n_device_buf.GetDeviceBuffer()},
|
||||
e_m_n_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
std::array<ck::index_t, 2>{StrideD0, StrideD1},
|
||||
StrideE,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
if(!device_op.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error("wrong! this device_op instance does not support this problem");
|
||||
}
|
||||
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
|
||||
sizeof(D0DataType) * N + sizeof(D1DataType) * M * N +
|
||||
sizeof(EDataType) * M * N;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
|
||||
float gb_per_sec = num_btype / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
|
||||
<< device_op.GetTypeString() << std::endl;
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
Tensor<AccDataType> c_m_n(HostTensorDescriptor(
|
||||
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
AccDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
PassThrough>;
|
||||
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument =
|
||||
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
for(int 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), d0_m_n(m, n), d1_m_n(m, n));
|
||||
}
|
||||
}
|
||||
|
||||
e_m_n_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData) ? 0 : 1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
@@ -1 +0,0 @@
|
||||
add_example_executable(example_gemm_xdl_bias_relu_add gemm_xdl_bias_relu_add.cpp)
|
||||
@@ -1,28 +0,0 @@
|
||||
# Instructions for ```example_gemm_xdl_bias_relu_add```
|
||||
|
||||
## Run ```example_gemm_xdl_bias_relu_add```
|
||||
```bash
|
||||
#arg1: verification (0=no, 1=yes)
|
||||
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
|
||||
#arg3: run kernel # of times (>1)
|
||||
#arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC
|
||||
./bin/example_gemm_xdl_bias_relu_add 0 1 5 3840 4096 4096 4096 4096 4096
|
||||
```
|
||||
|
||||
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
|
||||
```
|
||||
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
|
||||
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
|
||||
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
|
||||
c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
|
||||
c1_m_n: dim 2, lengths {3840, 4096}, strides {1, 0}
|
||||
arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
|
||||
arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
|
||||
arg.c_grid_desc_m_n_{ 3840, 4096}
|
||||
arg.c0_grid_desc_m_n_{ 3840, 4096}
|
||||
arg.c1_grid_desc_m_n_{ 3840, 4096}
|
||||
launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
|
||||
Warm up
|
||||
Start running 5 times...
|
||||
Perf: 1.27583 ms, 100.992 TFlops, 73.9688 GB/s
|
||||
```
|
||||
@@ -1,257 +0,0 @@
|
||||
#include <iostream>
|
||||
#include <numeric>
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
#include <stdlib.h>
|
||||
#include <half.hpp>
|
||||
|
||||
#include "check_err.hpp"
|
||||
#include "config.hpp"
|
||||
#include "print.hpp"
|
||||
#include "device.hpp"
|
||||
#include "host_tensor.hpp"
|
||||
#include "host_tensor_generator.hpp"
|
||||
#include "host_gemm.hpp"
|
||||
#include "device_tensor.hpp"
|
||||
#include "element_wise_operation.hpp"
|
||||
#include "device_gemm_xdl_c_shuffle_bias_activation_add.hpp"
|
||||
#include "reference_gemm_bias_activation_add.hpp"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using ADataType = ck::half_t;
|
||||
using BDataType = ck::half_t;
|
||||
using CDataType = ck::half_t;
|
||||
using AccDataType = float;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using CLayout = ck::tensor_layout::gemm::RowMajor;
|
||||
|
||||
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
|
||||
using CElementOp = ck::tensor_operation::element_wise::AddReluAdd;
|
||||
|
||||
// clang-format off
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle_Bias_Activation_Add<
|
||||
ADataType, // ADataType
|
||||
BDataType, // BDataType
|
||||
CDataType, // CDataType
|
||||
AccDataType, // AccDataType
|
||||
ALayout, // ALayout
|
||||
BLayout, // BLayout
|
||||
CLayout, // CLayout
|
||||
AElementOp, // AElementwiseOperation
|
||||
BElementOp, // BElementwiseOperation
|
||||
CElementOp, // CElementwiseOperation
|
||||
256, // BlockSize
|
||||
256, // MPerBlock
|
||||
128, // NPerBlock
|
||||
4, // K0PerBlock
|
||||
8, // K1
|
||||
32, // MPerXDL
|
||||
32, // NPerXDL
|
||||
4, // MXdlPerWave
|
||||
2, // NXdlPerWave
|
||||
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
|
||||
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
|
||||
2, // ABlockTransferSrcVectorDim
|
||||
8, // ABlockTransferSrcScalarPerVector
|
||||
8, // ABlockTransferDstScalarPerVector_K1
|
||||
true, // ABlockLdsAddExtraM
|
||||
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
|
||||
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
|
||||
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::ReferenceGemmBiasActivationAdd<ADataType,
|
||||
BDataType,
|
||||
CDataType,
|
||||
AElementOp,
|
||||
BElementOp,
|
||||
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 = 3840;
|
||||
ck::index_t N = 4096;
|
||||
ck::index_t K = 4096;
|
||||
|
||||
ck::index_t StrideA = 4096;
|
||||
ck::index_t StrideB = 4096;
|
||||
ck::index_t StrideC = 4096;
|
||||
ck::index_t StrideC1 = 4096;
|
||||
|
||||
if(argc == 4)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
}
|
||||
else if(argc == 11)
|
||||
{
|
||||
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]);
|
||||
|
||||
StrideA = std::stoi(argv[7]);
|
||||
StrideB = std::stoi(argv[8]);
|
||||
StrideC = std::stoi(argv[9]);
|
||||
StrideC1 = std::stoi(argv[10]);
|
||||
}
|
||||
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=n0, 1=yes)\n");
|
||||
printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, StrideC1\n");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
auto f_host_tensor_descriptor =
|
||||
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
||||
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
|
||||
std::vector<std::size_t>({stride, 1}));
|
||||
}
|
||||
else
|
||||
{
|
||||
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
|
||||
std::vector<std::size_t>({1, stride}));
|
||||
}
|
||||
};
|
||||
|
||||
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<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{}));
|
||||
|
||||
// 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})));
|
||||
|
||||
// c1_m_n[m ,n]
|
||||
Tensor<CDataType> c1_m_n(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
|
||||
std::cout << "a_m_k: " << a_m_k.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 << "c0_n: " << c0_n.mDesc << std::endl;
|
||||
std::cout << "c1_m_n: " << c1_m_n.mDesc << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
|
||||
c0_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
|
||||
c1_m_n.GenerateTensorValue(GeneratorTensor_2<CDataType>{-5, 5});
|
||||
break;
|
||||
default:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
c0_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
|
||||
c1_m_n.GenerateTensorValue(GeneratorTensor_3<CDataType>{0.0, 1.0});
|
||||
}
|
||||
|
||||
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
|
||||
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
|
||||
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
|
||||
DeviceMem c0_n_device_buf(sizeof(CDataType) * c0_n.mDesc.GetElementSpace());
|
||||
DeviceMem c1_m_n_device_buf(sizeof(CDataType) * c1_m_n.mDesc.GetElementSpace());
|
||||
|
||||
a_m_k_device_buf.ToDevice(a_m_k.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());
|
||||
c0_n_device_buf.ToDevice(c0_n.mData.data());
|
||||
c1_m_n_device_buf.ToDevice(c1_m_n.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto c_element_op = CElementOp{};
|
||||
|
||||
// do GEMM
|
||||
auto gemm = DeviceGemmInstance{};
|
||||
|
||||
auto invoker = gemm.MakeInvoker();
|
||||
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
|
||||
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c0_n_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c1_m_n_device_buf.GetDeviceBuffer()),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
StrideC1,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op);
|
||||
|
||||
if(!gemm.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
}
|
||||
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
|
||||
sizeof(CDataType) * M * N + sizeof(CDataType) * N +
|
||||
sizeof(CDataType) * M * N;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
|
||||
float gb_per_sec = num_btype / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
|
||||
<< std::endl;
|
||||
|
||||
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(a_m_k,
|
||||
b_k_n,
|
||||
c_m_n_host_result,
|
||||
c0_n,
|
||||
c1_m_n,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op);
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
@@ -14,7 +14,6 @@
|
||||
#include "element_wise_operation.hpp"
|
||||
#include "reference_gemm.hpp"
|
||||
#include "gemm_specialization.hpp"
|
||||
#include "element_wise_reduce_operation.hpp"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
@@ -14,7 +14,6 @@
|
||||
#include "element_wise_operation.hpp"
|
||||
#include "reference_gemm.hpp"
|
||||
#include "gemm_specialization.hpp"
|
||||
#include "element_wise_reduce_operation.hpp"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
@@ -14,7 +14,6 @@
|
||||
#include "element_wise_operation.hpp"
|
||||
#include "reference_gemm.hpp"
|
||||
#include "gemm_specialization.hpp"
|
||||
#include "element_wise_reduce_operation.hpp"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
@@ -39,7 +39,7 @@ endfunction(add_example_executable_no_testing EXAMPLE_NAME)
|
||||
add_subdirectory(01_gemm)
|
||||
add_subdirectory(02_gemm_alpha_beta)
|
||||
add_subdirectory(03_gemm_bias_relu)
|
||||
add_subdirectory(04_gemm_bias_relu_add)
|
||||
add_subdirectory(04_gemm_add_add_fastgelu)
|
||||
add_subdirectory(06_conv2d_fwd_bias_relu)
|
||||
add_subdirectory(07_conv2d_fwd_bias_relu_add)
|
||||
add_subdirectory(09_convnd_fwd)
|
||||
|
||||
Reference in New Issue
Block a user