mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-13 09:45:56 +00:00
500fa99512
* convnd_fwd fp16 example * update example * update example * update instance * updating refernce conv * update reference conv * update conv fwd profiler * update conv 1d and 3d instance * update include path * clean * update profiler for conv bwd data and weight * update conv bwd weight * clean * update conv example * update profiler for conv bwd weight * update ckprofiler for conv bwd data * fix reference conv bwd data bug; update conv bwd data test * update examples * fix initialization issue * update test for conv fwd * clean * clean * remove test case too sensitive to error threshhold * fix test * clean * fix build * adding conv multiple d * adding conv multiple D * add matrix padder * add gemm padding to convnd * adding group conv * update gemm multi-d * refactor * refactor * refactor * clean * clean * refactor * refactor * reorg * add ds * add bias * clean * add G * adding group * adding group * adding group * update Tensor * clean * update example * update DeviceGemmMultipleD_Xdl_CShuffle * update conv bwd-data and bwd-weight * upate contraction example * update gemm and batch gemm with e permute * fix example build * instance for grouped conv1d * update example * adding group conv instance * update gemm bilinear instance * update gemm+add+add+fastgelu instance * update profiler * update profiler * update test * update test and client example * clean * add grouped conv into profiler * update profiler * clean * add test grouped conv, update all conv test to gtest * update test
253 lines
12 KiB
C++
253 lines
12 KiB
C++
// SPDX-License-Identifier: MIT
|
|
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
|
|
|
#include <iostream>
|
|
#include <numeric>
|
|
#include <initializer_list>
|
|
#include <cstdlib>
|
|
|
|
#include "ck/ck.hpp"
|
|
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
|
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
|
|
#include "ck/tensor_operation/gpu/device/device_grouped_gemm_xdl.hpp"
|
|
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
|
|
|
#include "ck/library/utility/check_err.hpp"
|
|
#include "ck/library/utility/device_memory.hpp"
|
|
#include "ck/library/utility/host_tensor.hpp"
|
|
#include "ck/library/utility/host_tensor_generator.hpp"
|
|
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
|
|
|
template <ck::index_t... Is>
|
|
using S = ck::Sequence<Is...>;
|
|
|
|
using F16 = ck::half_t;
|
|
using F32 = float;
|
|
|
|
using Row = ck::tensor_layout::gemm::RowMajor;
|
|
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
|
|
|
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
|
|
|
using ADataType = F16;
|
|
using BDataType = F16;
|
|
using AccDataType = F32;
|
|
using CShuffleDataType = F16;
|
|
using DsDataType = ck::Tuple<>;
|
|
using EDataType = F16;
|
|
|
|
using ALayout = Row;
|
|
using BLayout = Col;
|
|
using DsLayout = ck::Tuple<>;
|
|
using ELayout = Row;
|
|
|
|
using AElementOp = PassThrough;
|
|
using BElementOp = PassThrough;
|
|
using CDEElementOp = PassThrough;
|
|
|
|
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
|
|
|
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
|
|
// clang-format off
|
|
//######| ALayout| BLayout| DsLayout| 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, DsLayout, 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
|
|
|
|
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
|
BDataType,
|
|
EDataType,
|
|
AccDataType,
|
|
AElementOp,
|
|
BElementOp,
|
|
CDEElementOp>;
|
|
|
|
int main(int argc, char* argv[])
|
|
{
|
|
bool do_verification = true;
|
|
int init_method = 1;
|
|
bool time_kernel = false;
|
|
|
|
if(argc == 4)
|
|
{
|
|
do_verification = std::stoi(argv[1]);
|
|
init_method = std::stoi(argv[2]);
|
|
time_kernel = std::stoi(argv[3]);
|
|
}
|
|
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");
|
|
exit(0);
|
|
}
|
|
|
|
int group_count = rand() % 16 + 1;
|
|
|
|
// GEMM shape
|
|
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
|
|
std::vector<const void*> p_a, p_b;
|
|
std::vector<void*> p_c;
|
|
|
|
gemm_descs.reserve(group_count);
|
|
|
|
for(int i = 0; i < group_count; i++)
|
|
{
|
|
int M = 256 + 256 * i;
|
|
int N = 128 + 128 * i;
|
|
int K = 64 + 64 * i;
|
|
|
|
int stride_A = K;
|
|
int stride_B = K;
|
|
int stride_C = N;
|
|
|
|
gemm_descs.push_back({M, N, K, stride_A, stride_B, stride_C, {}});
|
|
}
|
|
|
|
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}));
|
|
}
|
|
};
|
|
|
|
std::vector<Tensor<ADataType>> a_tensors;
|
|
std::vector<Tensor<BDataType>> b_tensors;
|
|
std::vector<Tensor<EDataType>> c_host_tensors;
|
|
std::vector<Tensor<EDataType>> c_device_tensors;
|
|
|
|
a_tensors.reserve(group_count);
|
|
b_tensors.reserve(group_count);
|
|
c_host_tensors.reserve(group_count);
|
|
c_device_tensors.reserve(group_count);
|
|
|
|
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
|
|
|
|
std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
|
|
|
|
a_tensors_device.reserve(group_count);
|
|
b_tensors_device.reserve(group_count);
|
|
c_tensors_device.reserve(group_count);
|
|
|
|
std::size_t flop = 0, num_btype = 0;
|
|
|
|
for(std::size_t i = 0; i < gemm_descs.size(); i++)
|
|
{
|
|
a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
|
|
gemm_descs[i].M_, gemm_descs[i].K_, gemm_descs[i].stride_A_, ALayout{})));
|
|
b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
|
|
gemm_descs[i].K_, gemm_descs[i].N_, gemm_descs[i].stride_B_, BLayout{})));
|
|
c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
|
|
gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
|
|
c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
|
|
gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
|
|
|
|
std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
|
|
<< " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
|
|
<< std::endl;
|
|
|
|
flop += std::size_t(2) * gemm_descs[i].M_ * gemm_descs[i].K_ * gemm_descs[i].N_;
|
|
num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
|
|
sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
|
|
sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
|
|
|
|
switch(init_method)
|
|
{
|
|
case 0: break;
|
|
case 1:
|
|
a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
|
|
b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
|
|
break;
|
|
case 2:
|
|
a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
|
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
|
break;
|
|
default:
|
|
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
|
|
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
|
|
}
|
|
}
|
|
|
|
for(std::size_t i = 0; i < gemm_descs.size(); i++)
|
|
{
|
|
a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
|
|
sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpaceSize()));
|
|
b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
|
|
sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpaceSize()));
|
|
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
|
|
sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSpaceSize()));
|
|
|
|
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
|
|
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
|
|
|
|
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
|
|
p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
|
|
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
|
|
}
|
|
|
|
auto a_element_op = AElementOp{};
|
|
auto b_element_op = BElementOp{};
|
|
auto c_element_op = CDEElementOp{};
|
|
|
|
auto gemm = DeviceGemmInstance{};
|
|
auto invoker = gemm.MakeInvoker();
|
|
|
|
std::vector<std::array<const void*, 0>> p_Ds = {};
|
|
|
|
// do GEMM
|
|
auto argument = gemm.MakeArgument(
|
|
p_a, p_b, p_Ds, p_c, gemm_descs, a_element_op, b_element_op, c_element_op);
|
|
|
|
DeviceMem gemm_desc_workspace(gemm.GetWorkSpaceSize(&argument));
|
|
|
|
gemm.SetWorkSpacePointer(&argument, gemm_desc_workspace.GetDeviceBuffer());
|
|
|
|
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});
|
|
|
|
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)
|
|
{
|
|
for(std::size_t i = 0; i < gemm_descs.size(); i++)
|
|
{
|
|
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
|
|
auto ref_gemm = ReferenceGemmInstance{};
|
|
auto ref_invoker = ref_gemm.MakeInvoker();
|
|
|
|
auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
|
|
b_tensors[i],
|
|
c_host_tensors[i],
|
|
a_element_op,
|
|
b_element_op,
|
|
c_element_op);
|
|
|
|
ref_invoker.Run(ref_argument);
|
|
pass &= ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
|
|
}
|
|
}
|
|
|
|
return pass ? 0 : 1;
|
|
}
|