mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-14 10:09:41 +00:00
Gemm reduce max (#209)
* [What] Rename the example
[Why] Prepare to add unary reduction
* Add global oparation to the parameter
* Add atomicmax
* Fix compile error
* Support atomicMax (hip library)
* Rename the reduction example
* Fix target name
* use p_d1_grid as the indicator directly
* Prevent performance issue. Let passthrough handle it.
* Implement the function template the specialize the float2
* No need to separate into two lines
* Remove empty line
* add comment
* Fix compile error due to merge from develop
* make the implementation of atomic_max / atomic_add explicit for each datatype
* Refine typo
* For future CI test
* Fix compiler error in ckProfiler
* Merge commit 'de2769e3a6695b38a20529261273ddc5cdaab2fe'
* simply use remove_pointer
* Rename type and var
* Refine example
* Modify reducemax example
* Fix bug in reduction
* Change initialize range
* Implement F64 version of atomicMax
* Move reduction code together
* Add buffer atomic_max
* Fix coding style by clang-format
* Integrate new api of DeviceGemmReduce_Xdl_CShuffle
* Integrate Batch gemm reduction
* Fix example
* fix example
* clean up
* Fix batch gemm tensor operation
* Fix coding style
* Fix template augument
* Fix clang format
* Keep flexible of different stride for each D tensor
* Fix compile error for ckProfiler
* Fix typo
* [What] Fix naming
[Why] Prepare to add out elementop
* Add DoutElementOp
Co-authored-by: Chao Liu <chao.liu2@amd.com>
Co-authored-by: rocking <chunylai@amd.com>
[ROCm/composable_kernel commit: 0ffe956ab1]
This commit is contained in:
rocking5566
@@ -1 +1,2 @@
|
||||
add_example_executable(example_gemm_reduce_xdl_fp16 gemm_reduce_xdl_fp16.cpp)
|
||||
add_example_executable(example_gemm_reduce_xdl_max_fp16 gemm_reduce_xdl_max_fp16.cpp)
|
||||
add_example_executable(example_gemm_reduce_xdl_sum_squaresum_fp16 gemm_reduce_xdl_sum_squaresum_fp16.cpp)
|
||||
|
||||
249
example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp
Normal file
249
example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp
Normal file
@@ -0,0 +1,249 @@
|
||||
#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 "device_gemm_reduce_xdl_cshuffle.hpp"
|
||||
#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...>;
|
||||
|
||||
using F16 = ck::half_t;
|
||||
using F32 = float;
|
||||
using F64 = double;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using CDataType = F16;
|
||||
using ReduceAccDataType = F32;
|
||||
using DDataType = F64;
|
||||
using DPtrsGlobal = ck::Tuple<DDataType*>;
|
||||
|
||||
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::PassThrough;
|
||||
using DsReduceOp = ck::Tuple<ck::reduce::Max<ReduceAccDataType>>;
|
||||
using DsElementOp = ck::Tuple<
|
||||
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>>;
|
||||
using DGlobalMemOp =
|
||||
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
|
||||
|
||||
static constexpr auto GemmSpecialization =
|
||||
ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
// clang-format off
|
||||
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
|
||||
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
|
||||
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
|
||||
//######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< Row, Col, Row, F16, F16, F16, F32, F32, ReduceAccDataType, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DsReduceOp, DsElementOp, DsElementOp, DGlobalMemOp, GemmSpecialization, 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, S<64, 4>, 4, 1>;
|
||||
// clang-format on
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::
|
||||
ReferenceGemm<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;
|
||||
|
||||
if(argc == 1)
|
||||
{
|
||||
// do nothing
|
||||
}
|
||||
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 == 10)
|
||||
{
|
||||
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]);
|
||||
}
|
||||
else
|
||||
{
|
||||
printf("arg1: verification (0=no, 1=yes)\n");
|
||||
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
|
||||
printf("arg3: run kernel # of times (>1)\n");
|
||||
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\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<DDataType> d_m_host_result(
|
||||
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
|
||||
|
||||
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
Tensor<DDataType> d_m_device_result(
|
||||
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
|
||||
|
||||
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 << "d_m: " << d_m_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});
|
||||
break;
|
||||
default:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
break;
|
||||
}
|
||||
|
||||
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
|
||||
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
|
||||
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
|
||||
DeviceMem d_device_buf(sizeof(DDataType) * d_m_device_result.mDesc.GetElementSpace());
|
||||
|
||||
a_device_buf.ToDevice(a_m_k.mData.data());
|
||||
b_device_buf.ToDevice(b_k_n.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto c_element_op = CElementOp{};
|
||||
auto ds_element_op = DsElementOp{};
|
||||
auto p_ds_global = ck::make_tuple(static_cast<DDataType*>(d_device_buf.GetDeviceBuffer()));
|
||||
|
||||
// do GEMM
|
||||
auto gemm = DeviceGemmReduceInstance{};
|
||||
auto invoker = gemm.MakeInvoker();
|
||||
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
|
||||
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
|
||||
p_ds_global,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op,
|
||||
ds_element_op,
|
||||
ds_element_op);
|
||||
|
||||
if(!gemm.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
}
|
||||
|
||||
// init D
|
||||
d_device_buf.SetValue(ck::NumericLimits<DDataType>::Lowest());
|
||||
|
||||
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(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, "
|
||||
<< gemm.GetTypeString() << std::endl;
|
||||
|
||||
bool pass = true;
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
|
||||
d_device_buf.FromDevice(d_m_device_result.mData.data());
|
||||
|
||||
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, a_element_op, b_element_op, c_element_op);
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
auto d_reduce_op = DsReduceOp{}[ck::Number<0>{}];
|
||||
|
||||
for(int m = 0; m < M; ++m)
|
||||
{
|
||||
ReduceAccDataType d_acc = d_reduce_op.GetReductionZeroVal();
|
||||
|
||||
for(int n = 0; n < N; ++n)
|
||||
d_reduce_op(d_acc, c_m_n_host_result(m, n));
|
||||
|
||||
d_m_host_result(m) = d_acc;
|
||||
}
|
||||
|
||||
pass = ck::utils::check_err(c_m_n_device_result.mData,
|
||||
c_m_n_host_result.mData,
|
||||
"Error: Incorrect results c") &&
|
||||
ck::utils::check_err(d_m_device_result.mData,
|
||||
d_m_host_result.mData,
|
||||
"Error: Incorrect results d",
|
||||
1e-3,
|
||||
1e-3);
|
||||
}
|
||||
|
||||
return pass ? 0 : 1;
|
||||
}
|
||||
@@ -3,7 +3,7 @@
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
#include <stdlib.h>
|
||||
#include <half.hpp>
|
||||
|
||||
#include "check_err.hpp"
|
||||
#include "config.hpp"
|
||||
#include "device.hpp"
|
||||
@@ -26,10 +26,12 @@ using F32 = float;
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using CDataType = F16;
|
||||
using DDataType = F32;
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using CDataType = F16;
|
||||
using ReduceAccDataType = F32;
|
||||
using DDataType = F32;
|
||||
using DPtrsGlobal = ck::Tuple<DDataType*, DDataType*>;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
@@ -38,20 +40,31 @@ 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::PassThrough;
|
||||
using D0ReduceOp = ck::reduce::Add<float>;
|
||||
using D1ReduceOp = ck::reduce::Add<float>;
|
||||
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
|
||||
using D0ReduceOp = ck::reduce::Add<ReduceAccDataType>;
|
||||
using D1ReduceOp = ck::reduce::Add<ReduceAccDataType>;
|
||||
using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
|
||||
|
||||
using UnaryIdenticElementOp =
|
||||
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>;
|
||||
using UnarySquareElementOp =
|
||||
ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
|
||||
using DxsInElementOp = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
|
||||
using DxsOutElementOp = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
|
||||
|
||||
using DGlobalMemOp =
|
||||
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
|
||||
ck::InMemoryDataOperationEnum::AtomicAdd>;
|
||||
|
||||
static constexpr auto GemmSpecialization =
|
||||
ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
// clang-format off
|
||||
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
|
||||
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
|
||||
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
|
||||
//######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, AElementOp, BElementOp, CElementOp, D0ReduceOp, D1ReduceOp, D1ElementOp, GemmSpecialization, 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, S<64, 4>, 4, 1>;
|
||||
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
|
||||
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
|
||||
//######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp, DGlobalMemOp, GemmSpecialization, 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, S<64, 4>, 4, 1>;
|
||||
// clang-format on
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::
|
||||
@@ -162,10 +175,11 @@ int main(int argc, char* argv[])
|
||||
a_device_buf.ToDevice(a_m_k.mData.data());
|
||||
b_device_buf.ToDevice(b_k_n.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto c_element_op = CElementOp{};
|
||||
auto d1_element_op = D1ElementOp{};
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto c_element_op = CElementOp{};
|
||||
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
|
||||
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
|
||||
|
||||
// do GEMM
|
||||
auto gemm = DeviceGemmReduceInstance{};
|
||||
@@ -173,8 +187,7 @@ int main(int argc, char* argv[])
|
||||
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
|
||||
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
|
||||
static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
|
||||
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
|
||||
dxs_global,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
@@ -184,7 +197,8 @@ int main(int argc, char* argv[])
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op,
|
||||
d1_element_op);
|
||||
DxsInElementOp{},
|
||||
DxsOutElementOp{});
|
||||
|
||||
if(!gemm.IsSupportedArgument(argument))
|
||||
{
|
||||
@@ -213,6 +227,7 @@ int main(int argc, char* argv[])
|
||||
<< gemm.GetTypeString() << std::endl;
|
||||
|
||||
bool pass = true;
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
|
||||
@@ -237,10 +252,12 @@ int main(int argc, char* argv[])
|
||||
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
float d0_val = ck::type_convert<float>(c_m_n_host_result(m, n));
|
||||
float d1_val;
|
||||
float c_val = ck::type_convert<float>(c_m_n_host_result(m, n));
|
||||
float d0_val = 0;
|
||||
float d1_val = 0;
|
||||
|
||||
d1_element_op(d1_val, d0_val);
|
||||
UnaryIdenticElementOp{}(d0_val, c_val);
|
||||
UnarySquareElementOp{}(d1_val, c_val);
|
||||
d0_reduce_op(d0_acc, d0_val);
|
||||
d1_reduce_op(d1_acc, d1_val);
|
||||
}
|
||||
@@ -249,18 +266,19 @@ int main(int argc, char* argv[])
|
||||
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
|
||||
}
|
||||
|
||||
pass &= ck::utils::check_err(
|
||||
c_m_n_device_result.mData, c_m_n_host_result.mData, "Error: Incorrect results c");
|
||||
pass &= ck::utils::check_err(d0_m_device_result.mData,
|
||||
d0_m_host_result.mData,
|
||||
"Error: Incorrect results d0",
|
||||
1e-3,
|
||||
1e-3);
|
||||
pass &= ck::utils::check_err(d1_m_device_result.mData,
|
||||
d1_m_host_result.mData,
|
||||
"Error: Incorrect results d1",
|
||||
1e-3,
|
||||
1e-3);
|
||||
pass = ck::utils::check_err(c_m_n_device_result.mData,
|
||||
c_m_n_host_result.mData,
|
||||
"Error: Incorrect results c") &&
|
||||
ck::utils::check_err(d0_m_device_result.mData,
|
||||
d0_m_host_result.mData,
|
||||
"Error: Incorrect results d0",
|
||||
1e-4,
|
||||
1e-5) &&
|
||||
ck::utils::check_err(d1_m_device_result.mData,
|
||||
d1_m_host_result.mData,
|
||||
"Error: Incorrect results d1",
|
||||
1e-3,
|
||||
1e-5);
|
||||
}
|
||||
|
||||
return pass ? 0 : 1;
|
||||
@@ -25,10 +25,12 @@ using F32 = float;
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using CDataType = F16;
|
||||
using DDataType = F32;
|
||||
using ADataType = F16;
|
||||
using BDataType = F16;
|
||||
using CDataType = F16;
|
||||
using ReduceAccDataType = F32;
|
||||
using DDataType = F32;
|
||||
using DPtrsGlobal = ck::Tuple<DDataType*, DDataType*>;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
@@ -37,20 +39,31 @@ 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::PassThrough;
|
||||
using D0ReduceOp = ck::reduce::Add<float>;
|
||||
using D1ReduceOp = ck::reduce::Add<float>;
|
||||
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
|
||||
using D0ReduceOp = ck::reduce::Add<ReduceAccDataType>;
|
||||
using D1ReduceOp = ck::reduce::Add<ReduceAccDataType>;
|
||||
using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
|
||||
|
||||
using UnaryIdenticElementOp =
|
||||
ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>;
|
||||
using UnarySquareElementOp =
|
||||
ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
|
||||
using DxsInElementOp = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
|
||||
using DxsOutElementOp = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
|
||||
|
||||
using DGlobalMemOp =
|
||||
ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
|
||||
ck::InMemoryDataOperationEnum::AtomicAdd>;
|
||||
|
||||
static constexpr auto GemmSpecialization =
|
||||
ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
// clang-format off
|
||||
using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatchedGemmReduce_Xdl_CShuffle
|
||||
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
|
||||
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
|
||||
//######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, AElementOp, BElementOp, CElementOp, D0ReduceOp, D1ReduceOp, D1ElementOp, GemmSpecialization, 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, S<64, 4>, 4, 1>;
|
||||
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
|
||||
//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
|
||||
//######| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
|
||||
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp, DGlobalMemOp, GemmSpecialization, 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, S<64, 4>, 4, 1>;
|
||||
// clang-format on
|
||||
|
||||
using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
|
||||
@@ -170,12 +183,11 @@ int main(int argc, char* argv[])
|
||||
a_device_buf.ToDevice(a_g_m_k.mData.data());
|
||||
b_device_buf.ToDevice(b_g_k_n.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto c_element_op = CElementOp{};
|
||||
auto d0_reduce_op = D0ReduceOp{};
|
||||
auto d1_reduce_op = D1ReduceOp{};
|
||||
auto d1_element_op = D1ElementOp{};
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto c_element_op = CElementOp{};
|
||||
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
|
||||
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
|
||||
|
||||
// do GEMM
|
||||
auto batched_gemm = DeviceBatchedGemmReduceInstance{};
|
||||
@@ -184,8 +196,7 @@ int main(int argc, char* argv[])
|
||||
batched_gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
|
||||
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
|
||||
static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
|
||||
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
|
||||
dxs_global,
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
@@ -195,7 +206,8 @@ int main(int argc, char* argv[])
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op,
|
||||
d1_element_op,
|
||||
DxsInElementOp{},
|
||||
DxsOutElementOp{},
|
||||
BatchCount);
|
||||
|
||||
if(!batched_gemm.IsSupportedArgument(argument))
|
||||
@@ -240,6 +252,9 @@ int main(int argc, char* argv[])
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
auto d0_reduce_op = D0ReduceOp{};
|
||||
auto d1_reduce_op = D1ReduceOp{};
|
||||
|
||||
for(int batch = 0; batch < BatchCount; ++batch)
|
||||
{
|
||||
for(int m = 0; m < M; ++m)
|
||||
@@ -249,10 +264,12 @@ int main(int argc, char* argv[])
|
||||
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
float d0_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
|
||||
float d1_val;
|
||||
float c_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
|
||||
float d0_val = 0;
|
||||
float d1_val = 0;
|
||||
|
||||
d1_element_op(d1_val, d0_val);
|
||||
UnaryIdenticElementOp{}(d0_val, c_val);
|
||||
UnarySquareElementOp{}(d1_val, c_val);
|
||||
d0_reduce_op(d0_acc, d0_val);
|
||||
d1_reduce_op(d1_acc, d1_val);
|
||||
}
|
||||
@@ -262,17 +279,19 @@ int main(int argc, char* argv[])
|
||||
}
|
||||
}
|
||||
|
||||
pass &= ck::utils::check_err(c_g_m_n_host_result.mData, c_g_m_n_device_result.mData);
|
||||
pass &= ck::utils::check_err(d0_g_m_device_result.mData,
|
||||
d0_g_m_host_result.mData,
|
||||
"Error: Incorrect results! D0",
|
||||
1e-3,
|
||||
1e-3);
|
||||
pass &= ck::utils::check_err(d1_g_m_device_result.mData,
|
||||
d1_g_m_host_result.mData,
|
||||
"Error: Incorrect results! D1",
|
||||
1e-3,
|
||||
1e-3);
|
||||
pass = ck::utils::check_err(c_g_m_n_host_result.mData,
|
||||
c_g_m_n_device_result.mData,
|
||||
"Error: Incorrect results c") &&
|
||||
ck::utils::check_err(d0_g_m_device_result.mData,
|
||||
d0_g_m_host_result.mData,
|
||||
"Error: Incorrect results! D0",
|
||||
1e-4,
|
||||
1e-5) &&
|
||||
ck::utils::check_err(d1_g_m_device_result.mData,
|
||||
d1_g_m_host_result.mData,
|
||||
"Error: Incorrect results! D1",
|
||||
1e-3,
|
||||
1e-5);
|
||||
}
|
||||
|
||||
return pass ? 0 : 1;
|
||||
|
||||
@@ -33,7 +33,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
|
||||
add_executable(${EXAMPLE_NAME} ${FILE_NAME})
|
||||
target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor)
|
||||
add_dependencies(examples ${EXAMPLE_NAME})
|
||||
endfunction(add_example_executable EXAMPLE_NAME)
|
||||
endfunction(add_example_executable_no_testing EXAMPLE_NAME)
|
||||
|
||||
add_subdirectory(01_gemm)
|
||||
add_subdirectory(02_gemm_alpha_beta)
|
||||
|
||||
Reference in New Issue
Block a user