mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
Added Multi_ABD support into Gemm and GroupedGemmFixedNK (#978)
* added an example grouped_gemm_multi_abd * fixed ci * add setElementwiseOp * changed API * clean code: add multiA into example * fixed v7r2 copy * add transpose * clean * fixed vector_load check * Update example/15_grouped_gemm/grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update example/15_grouped_gemm/grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update example/15_grouped_gemm/grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update include/ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update include/ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update include/ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update include/ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update include/ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update include/ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update include/ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd_fixed_nk.hpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update include/ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd_fixed_nk.hpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * add reduce * testing * add example_b16_i8 * refactor example * clean * add mpading * disable reduce for kbatch = 1 * seperate reduce device op * add reduce op * add guard for workspace_size * add instances * format * fixed * add client example * add a colmajor * add instances * Update cmake-ck-dev.sh * Update profile_gemm_splitk.cpp * Update gridwise_gemm_xdlops_v2r4r2.hpp * format * Update profile_gemm_splitk.cpp * fixed * fixed * adjust test * adjust precision loss * adjust test * fixed * add bf16_i8 scale bias * fixed scale * fixed scale elementwise_op * revert contraction deviceop changes * fixed * Add AddFastGelu * Revert "Merge branch 'jizhan/gemm_splitk_reduce' into grouped_gemm_multi_abd_fixed_nk_example" This reverts commit3b5d001efd, reversing changes made to943199a991. * add Scales into elementwise * add gemm_multi_abd client example * add client examples * add rcr and crr * add grouped gemm client example * add grouped gemm client example * add instance for rcr crr * format * fixed * fixed cmake * fixed * fixed client_example * format * fixed contraction isSupport * Update include/ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd_fixed_nk.hpp Co-authored-by: Bartłomiej Kocot <barkocot@amd.com> * Update device_reduce_threadwise.hpp * clean * Fixes * Fix example --------- Co-authored-by: Jing Zhang <jizha@amd.com> Co-authored-by: Bartłomiej Kocot <barkocot@amd.com>
This commit is contained in:
zjing14
@@ -1 +1,2 @@
|
||||
add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
|
||||
add_example_executable(example_gemm_multi_ABD_xdl_bf16_i8 gemm_multi_ABD_xdl_bf16_i8.cpp)
|
||||
270
example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_bf16_i8.cpp
Normal file
270
example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_bf16_i8.cpp
Normal file
@@ -0,0 +1,270 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, 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/gemm_specialization.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/utility/device_memory.hpp"
|
||||
#include "ck/library/utility/host_tensor.hpp"
|
||||
#include "ck/library/utility/host_tensor_generator.hpp"
|
||||
#include "ck/library/utility/literals.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
|
||||
#include "ck/library/utility/check_err.hpp"
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using BF16 = ck::bhalf_t;
|
||||
using I8 = int8_t;
|
||||
using F32 = float;
|
||||
|
||||
using Row = ck::tensor_layout::gemm::RowMajor;
|
||||
using Col = ck::tensor_layout::gemm::ColumnMajor;
|
||||
|
||||
using A0DataType = BF16;
|
||||
using AsDataType = ck::Tuple<A0DataType>;
|
||||
using B0DataType = I8;
|
||||
using B1DataType = BF16;
|
||||
using BsDataType = ck::Tuple<B0DataType, B1DataType>;
|
||||
using AccDataType = F32;
|
||||
using CShuffleDataType = BF16;
|
||||
using D0DataType = BF16;
|
||||
using DsDataType = ck::Tuple<D0DataType>;
|
||||
using EDataType = BF16;
|
||||
|
||||
using A0Layout = Row;
|
||||
using AsLayout = ck::Tuple<A0Layout>;
|
||||
using B0Layout = Col;
|
||||
using B1Layout = B0Layout;
|
||||
using BsLayout = ck::Tuple<B0Layout, B1Layout>;
|
||||
using D0Layout = Row;
|
||||
using DsLayout = ck::Tuple<D0Layout>;
|
||||
using ELayout = Row;
|
||||
|
||||
using Scales = ck::tensor_operation::element_wise::Scales;
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
|
||||
|
||||
using AElementOp = PassThrough;
|
||||
using BElementOp = Scales;
|
||||
using CDEElementOp = AddFastGelu;
|
||||
|
||||
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
|
||||
// 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|
|
||||
///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
|
||||
< AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 128, 16, 128, 32, 8, 8, 16, 16, 1, 4, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 1, 1, 1, 1, S<1, 16, 1, 8>, 1>;
|
||||
|
||||
// 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 = 64;
|
||||
ck::index_t N = 1024;
|
||||
ck::index_t K = 512;
|
||||
|
||||
ck::index_t StrideA = K;
|
||||
ck::index_t StrideB = K;
|
||||
ck::index_t StrideD = N;
|
||||
ck::index_t StrideE = N;
|
||||
|
||||
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 == 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]);
|
||||
StrideD = std::stoi(argv[9]);
|
||||
StrideE = 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=no, 1=yes)\n");
|
||||
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
auto f_host_tensor_descriptor =
|
||||
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
||||
using namespace ck::literals;
|
||||
|
||||
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {stride, 1_uz});
|
||||
}
|
||||
else
|
||||
{
|
||||
return HostTensorDescriptor({row, col}, {1_uz, stride});
|
||||
}
|
||||
};
|
||||
|
||||
Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
|
||||
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
|
||||
Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
|
||||
Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
|
||||
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 << "a0_m_k: " << a0_m_k.mDesc << std::endl;
|
||||
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
|
||||
std::cout << "b1_k_n: " << b1_k_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)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
|
||||
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
|
||||
d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
|
||||
break;
|
||||
default:
|
||||
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
|
||||
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
|
||||
b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
|
||||
d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
|
||||
}
|
||||
|
||||
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
a0_device_buf.ToDevice(a0_m_k.mData.data());
|
||||
b0_device_buf.ToDevice(b0_k_n.mData.data());
|
||||
b1_device_buf.ToDevice(b1_k_n.mData.data());
|
||||
d_device_buf.ToDevice(d_m_n.mData.data());
|
||||
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
auto a_element_op = AElementOp{};
|
||||
auto b_element_op = BElementOp{};
|
||||
auto cde_element_op = CDEElementOp{};
|
||||
|
||||
constexpr ck::index_t NumATensor = 1;
|
||||
constexpr ck::index_t NumBTensor = 2;
|
||||
constexpr ck::index_t NumDTensor = 1;
|
||||
|
||||
// do GEMM
|
||||
auto device_op = DeviceOpInstance{};
|
||||
auto invoker = device_op.MakeInvoker();
|
||||
auto argument =
|
||||
device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
|
||||
std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
|
||||
b1_device_buf.GetDeviceBuffer()},
|
||||
std::array<const void*, NumDTensor>{d_device_buf.GetDeviceBuffer()},
|
||||
e_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
std::array<ck::index_t, NumATensor>{StrideA},
|
||||
std::array<ck::index_t, NumBTensor>{StrideB, 0},
|
||||
std::array<ck::index_t, NumDTensor>{StrideD},
|
||||
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");
|
||||
}
|
||||
|
||||
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(A0DataType) * M * K + sizeof(B0DataType) * K * 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"
|
||||
<< std::endl;
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
Tensor<CShuffleDataType> c_m_n({M, N});
|
||||
|
||||
Tensor<A0DataType> a_m_k({M, K});
|
||||
|
||||
Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
|
||||
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
|
||||
}
|
||||
}
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
|
||||
B1DataType,
|
||||
CShuffleDataType,
|
||||
AccDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>;
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(
|
||||
a0_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, 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), d_m_n(m, n));
|
||||
}
|
||||
}
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
@@ -37,7 +37,7 @@ using DDataType = F16;
|
||||
using EDataType = F16;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Col;
|
||||
using BLayout = Row;
|
||||
using DLayout = Row;
|
||||
using ELayout = Row;
|
||||
|
||||
@@ -141,9 +141,9 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl
|
||||
S<4, 64, 1>,
|
||||
S<1, 0, 2>,
|
||||
S<1, 0, 2>,
|
||||
1,
|
||||
2,
|
||||
8,
|
||||
8,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
@@ -161,10 +161,10 @@ int main(int argc, char* argv[])
|
||||
ck::index_t N = 4096;
|
||||
ck::index_t K = 4096;
|
||||
|
||||
ck::index_t StrideA = 4096;
|
||||
ck::index_t StrideB = 4096;
|
||||
ck::index_t StrideD = 4096;
|
||||
ck::index_t StrideE = 4096;
|
||||
ck::index_t StrideA = K;
|
||||
ck::index_t StrideB = N;
|
||||
ck::index_t StrideD = N;
|
||||
ck::index_t StrideE = N;
|
||||
|
||||
float alpha = 1.0f;
|
||||
float beta = 1.0f;
|
||||
|
||||
Reference in New Issue
Block a user