Skip to content

GitLab

Commit 421b718d authored by rocking's avatar rocking
Browse files

Merge commit '57fadf6f' into gemm_layernorm_welford

parents 2cf6f30b 57fadf6f
Hide whitespace changes
Inline Side-by-side
Showing with 1079 additions and 448 deletions
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
View file @ 421b718d
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstddef> #include "common.hpp"
#include <iostream>
#include <stdexcept>
#include <string>
#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_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.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/utility/literals.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 ADataType = F16;
using BDataType = F16; using BDataType = F16;
...@@ -62,6 +34,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C ...@@ -62,6 +34,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C
< 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>; < 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 // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
#include "run_gemm_add_add_fastgelu_example.inc" #include "run_gemm_add_add_fastgelu_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); } int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
View file @ 421b718d
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstddef> #include "common.hpp"
#include <iostream>
#include <stdexcept>
#include <string>
#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_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.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/utility/literals.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 = F32; using ADataType = F32;
using BDataType = F32; using BDataType = F32;
...@@ -62,6 +34,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C ...@@ -62,6 +34,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C
< 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, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>; < 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, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>;
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
#include "run_gemm_add_add_fastgelu_example.inc" #include "run_gemm_add_add_fastgelu_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); } int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp 0 → 100644
View file @ 421b718d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
#error Should compile this file with ck::int4_t support
#endif
#include "common.hpp"
using ADataType = I4;
using BDataType = I4;
using AccDataType = I32;
using CShuffleDataType = I32;
using D0DataType = I4;
using D1DataType = I4;
using DsDataType = ck::Tuple<D0DataType, D1DataType>;
using EDataType = I4;
using KernelADataType = I8;
using KernelBDataType = I8;
using KernelD0DataType = I8;
using KernelD1DataType = I8;
using KernelDsDataType = ck::Tuple<KernelD0DataType, KernelD1DataType>;
using KernelEDataType = I8;
using ALayout = Row;
using BLayout = Col;
using D0Layout = Row;
using D1Layout = Row;
using DsLayout = ck::Tuple<D0Layout, D1Layout>;
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| 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, KernelADataType, KernelBDataType, AccDataType, CShuffleDataType, KernelDsDataType, KernelEDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 16>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
#define BUILD_INT4_EXAMPLE
#include "run_gemm_add_add_fastgelu_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
View file @ 421b718d
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstddef> #include "common.hpp"
#include <iostream>
#include <stdexcept>
#include <string>
#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_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.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/utility/literals.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using I8 = int8_t;
using I32 = int32_t;
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 = I8; using ADataType = I8;
using BDataType = I8; using BDataType = I8;
...@@ -62,6 +34,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C ...@@ -62,6 +34,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_C
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 16>; < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 16>;
// clang-format on // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
#include "run_gemm_add_add_fastgelu_example.inc" #include "run_gemm_add_add_fastgelu_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); } int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
View file @ 421b718d
#pragma once #pragma once
struct ProblemSize final
{
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;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionConfig& config) bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionConfig& config)
{ {
#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
#endif
using namespace ck::literals; using namespace ck::literals;
auto& [M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE] = problem_size; auto& [M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE] = problem_size;
...@@ -43,7 +26,14 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC ...@@ -43,7 +26,14 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{})); 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<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_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{})); Tensor<
#ifdef BUILD_INT4_EXAMPLE
KernelEDataType
#else
EDataType
#endif
>
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 << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
...@@ -73,10 +63,22 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC ...@@ -73,10 +63,22 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC
DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize()); DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize()); DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
#ifdef BUILD_INT4_EXAMPLE
const Tensor<KernelADataType> a_m_k_converted(a_m_k);
const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
const Tensor<KernelD0DataType> d0_m_n_converted(d0_m_n);
const Tensor<KernelD1DataType> d1_m_n_converted(d1_m_n);
a_device_buf.ToDevice(a_m_k_converted.mData.data());
b_device_buf.ToDevice(b_k_n_converted.mData.data());
d0_device_buf.ToDevice(d0_m_n_converted.mData.data());
d1_device_buf.ToDevice(d1_m_n_converted.mData.data());
#else
a_device_buf.ToDevice(a_m_k.mData.data()); a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data()); b_device_buf.ToDevice(b_k_n.mData.data());
d0_device_buf.ToDevice(d0_m_n.mData.data()); d0_device_buf.ToDevice(d0_m_n.mData.data());
d1_device_buf.ToDevice(d1_m_n.mData.data()); d1_device_buf.ToDevice(d1_m_n.mData.data());
#endif
auto a_element_op = AElementOp{}; auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{}; auto b_element_op = BElementOp{};
...@@ -124,14 +126,6 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC ...@@ -124,14 +126,6 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC
{ {
Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N}); Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
...@@ -150,7 +144,13 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC ...@@ -150,7 +144,13 @@ bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionC
e_device_buf.FromDevice(e_m_n_device_result.mData.data()); e_device_buf.FromDevice(e_m_n_device_result.mData.data());
#ifdef BUILD_INT4_EXAMPLE
const Tensor<EDataType> e_m_n_device_result_converted(e_m_n_device_result);
return ck::utils::check_err(e_m_n_device_result_converted.mData, e_m_n_host_result.mData);
#else
return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData); return ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
#endif
} }
return true; return true;
...@@ -161,43 +161,6 @@ bool run_gemm_add_add_fastgelu_example(int argc, char* argv[]) ...@@ -161,43 +161,6 @@ bool run_gemm_add_add_fastgelu_example(int argc, char* argv[])
ProblemSize problem_size; ProblemSize problem_size;
ExecutionConfig config; ExecutionConfig config;
if(argc == 1) return !parse_cmd_args(argc, argv, problem_size, config) ||
{ run_gemm_add_add_fastgelu(problem_size, config);
// use default case
}
else if(argc == 4)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
}
else if(argc == 12)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.time_kernel = std::stoi(argv[3]);
problem_size.M = std::stoi(argv[4]);
problem_size.N = std::stoi(argv[5]);
problem_size.K = std::stoi(argv[6]);
problem_size.StrideA = std::stoi(argv[7]);
problem_size.StrideB = std::stoi(argv[8]);
problem_size.StrideD0 = std::stoi(argv[9]);
problem_size.StrideD1 = std::stoi(argv[10]);
problem_size.StrideE = std::stoi(argv[11]);
}
else
{
std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
<< std::endl
<< "arg3: time kernel (0=no, 1=yes)" << std::endl
<< "arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, "
"StrideE"
<< std::endl;
return true;
}
return run_gemm_add_add_fastgelu(problem_size, config);
} }
example/12_reduce/reduce_blockwise.cpp
View file @ 421b718d
...@@ -225,6 +225,28 @@ int main(int argc, char* argv[]) ...@@ -225,6 +225,28 @@ int main(int argc, char* argv[])
arg.scales[0], arg.scales[0],
arg.scales[1]); arg.scales[1]);
} }
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
else if(arg.data_type == 7)
{
pass = reduce_blockwise_test<int4_t, int32_t, ReduceTensorOp::AVG, false, false>(
arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inLengths,
arg.reduceDims,
arg.scales[0],
arg.scales[1]);
pass = pass && reduce_blockwise_test<int4_t, int8_t, ReduceTensorOp::MAX, false, false>(
arg.do_verification,
arg.init_method,
arg.time_kernel,
arg.inLengths,
arg.reduceDims,
arg.scales[0],
arg.scales[1]);
}
#endif
} }
else else
{ {
...@@ -251,6 +273,15 @@ int main(int argc, char* argv[]) ...@@ -251,6 +273,15 @@ int main(int argc, char* argv[])
pass && reduce_blockwise_test<int8_t, int32_t, ReduceOpId, PropagateNan, OutputIndex>( pass && reduce_blockwise_test<int8_t, int32_t, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f); true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
// for testing int4_t using AVG operation
pass = pass && reduce_blockwise_test<int4_t, int32_t, ReduceTensorOp::AVG, false, false>(
true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
// for testing int4_t using MAX operation
pass = pass && reduce_blockwise_test<int4_t, int8_t, ReduceTensorOp::MAX, false, false>(
true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
#endif
// for testing 3D input // for testing 3D input
pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>( pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
true, 2, true, {16, 64, 960}, {0, 1}, 1.0f, 0.0f); true, 2, true, {16, 64, 960}, {0, 1}, 1.0f, 0.0f);
......
example/12_reduce/reduce_blockwise_impl.hpp
View file @ 421b718d
...@@ -58,28 +58,47 @@ int reduce_blockwise_impl(bool do_verification, ...@@ -58,28 +58,47 @@ int reduce_blockwise_impl(bool do_verification,
std::is_same<InOutDataType, float>::value && std::is_same<InOutDataType, float>::value &&
(op_support_indices && !std::is_same<AccDataType, float>::value); (op_support_indices && !std::is_same<AccDataType, float>::value);
// 1) If InOutDataType is int8_t, must use int8_t as AccDataType for indexable reduction // 1) If InOutDataType is int8_t or int4_t, must use int8_t as AccDataType for indexable
// operations 2) If InOutDataType is int8_t, must use int32_t as AccDataType for non-indexable // reduction operations 2) If InOutDataType is int8_t or int4_t, must use int32_t as AccDataType
// reduction operations // for non-indexable reduction operations
constexpr bool invalid_reduce_4 = constexpr bool invalid_reduce_4 =
std::is_same<InOutDataType, int8_t>::value && std::is_same<InOutDataType, int8_t>::value &&
((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) || ((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
(op_support_indices && !std::is_same<AccDataType, int8_t>::value)); (op_support_indices && !std::is_same<AccDataType, int8_t>::value));
// 1) If InOutDataType is int8_t, the supported operation must be either indexable operations or #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
// ADD/AVG constexpr bool invalid_reduce_4_2 =
std::is_same<InOutDataType, int4_t>::value &&
((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
(op_support_indices && !std::is_same<AccDataType, int8_t>::value));
#endif
// 1) If InOutDataType is int8_t or int4_t, the supported operation must be either indexable
// operations or ADD/AVG
constexpr bool invalid_reduce_5 = std::is_same<InOutDataType, int8_t>::value && constexpr bool invalid_reduce_5 = std::is_same<InOutDataType, int8_t>::value &&
(!op_support_indices && ReduceOpId != ReduceTensorOp::ADD && (!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
ReduceOpId != ReduceTensorOp::AVG); ReduceOpId != ReduceTensorOp::AVG);
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
constexpr bool invalid_reduce_5_2 = std::is_same<InOutDataType, int4_t>::value &&
(!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
ReduceOpId != ReduceTensorOp::AVG);
#endif
// 1) If InOutDataType is bhalf_t, must use float as AccDataType for all reduction operations // 1) If InOutDataType is bhalf_t, must use float as AccDataType for all reduction operations
constexpr bool invalid_reduce_6 = constexpr bool invalid_reduce_6 =
std::is_same<InOutDataType, bhalf_t>::value && !std::is_same<AccDataType, float>::value; std::is_same<InOutDataType, bhalf_t>::value && !std::is_same<AccDataType, float>::value;
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
constexpr bool invalid_reduce =
(invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 || invalid_reduce_4 ||
invalid_reduce_5 || invalid_reduce_6 || invalid_reduce_4_2 || invalid_reduce_5_2);
#else
constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 || constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6); invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
#endif
if(invalid_reduce) if constexpr(invalid_reduce)
{ {
std::cerr << "The reduction setting is invalid, exiting!" << std::endl; std::cerr << "The reduction setting is invalid, exiting!" << std::endl;
return (-1); return (-1);
...@@ -91,10 +110,17 @@ int reduce_blockwise_impl(bool do_verification, ...@@ -91,10 +110,17 @@ int reduce_blockwise_impl(bool do_verification,
using AccElementwiseOperation = using AccElementwiseOperation =
typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation; typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
using InOutDataTypeInDevice = typename std::
conditional<std::is_same<InOutDataType, int4_t>::value, int8_t, InOutDataType>::type;
#else
using InOutDataTypeInDevice = InOutDataType;
#endif
using DeviceReduceInstance = using DeviceReduceInstance =
ck::tensor_operation::device::DeviceReduceMultiBlock<InOutDataType, ck::tensor_operation::device::DeviceReduceMultiBlock<InOutDataTypeInDevice,
AccDataType, AccDataType,
InOutDataType, InOutDataTypeInDevice,
Rank, Rank,
NumReduceDim, NumReduceDim,
ReduceOperation, ReduceOperation,
...@@ -166,13 +192,35 @@ int reduce_blockwise_impl(bool do_verification, ...@@ -166,13 +192,35 @@ int reduce_blockwise_impl(bool do_verification,
}; };
// these buffers are usually provided by the user application // these buffers are usually provided by the user application
DeviceMem in_dev(sizeof(InOutDataType) * in.mDesc.GetElementSpaceSize()); DeviceMem in_dev(sizeof(InOutDataTypeInDevice) * in.mDesc.GetElementSpaceSize());
DeviceMem out_dev(sizeof(InOutDataType) * out.mDesc.GetElementSpaceSize()); DeviceMem out_dev(sizeof(InOutDataTypeInDevice) * out.mDesc.GetElementSpaceSize());
in_dev.ToDevice(in.mData.data()); #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
if(std::is_same<InOutDataType, int4_t>::value)
{
std::vector<InOutDataTypeInDevice> tmp_buf(in.mData.size());
std::copy_n(in.mData.data(), in.mData.size(), tmp_buf.data());
in_dev.ToDevice(tmp_buf.data());
}
else
#endif
in_dev.ToDevice(in.mData.data());
if(beta != 0.0f) if(beta != 0.0f)
out_dev.ToDevice(out.mData.data()); {
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
if(std::is_same<InOutDataType, int4_t>::value)
{
std::vector<InOutDataTypeInDevice> tmp_buf(in.mData.size());
std::copy_n(out.mData.data(), out.mData.size(), tmp_buf.data());
out_dev.ToDevice(tmp_buf.data());
}
else
#endif
out_dev.ToDevice(out.mData.data());
};
size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0; size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
...@@ -261,7 +309,19 @@ int reduce_blockwise_impl(bool do_verification, ...@@ -261,7 +309,19 @@ int reduce_blockwise_impl(bool do_verification,
if(do_verification) if(do_verification)
{ {
out_dev.FromDevice(out.mData.data()); #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
if(std::is_same<InOutDataType, int4_t>::value)
{
std::vector<InOutDataTypeInDevice> tmp_buf(out.mData.size());
out_dev.FromDevice(tmp_buf.data());
std::copy_n(tmp_buf.data(), out.mData.size(), out.mData.data());
}
else
#endif
out_dev.FromDevice(out.mData.data());
pass = pass && ck::utils::check_err(out.mData, out_ref.mData); pass = pass && ck::utils::check_err(out.mData, out_ref.mData);
if(OutputIndex) if(OutputIndex)
......
example/15_grouped_gemm/CMakeLists.txt
View file @ 421b718d
add_custom_target(example_grouped_gemm_xdl)
add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp) add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
add_dependencies(example_grouped_gemm_xdl
example_grouped_gemm_xdl_fp32
example_grouped_gemm_xdl_fp16
example_grouped_gemm_xdl_bfp16
example_grouped_gemm_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
endif()
example/15_grouped_gemm/grouped_gemm_xdl_bfp16.cpp 0 → 100644
View file @ 421b718d
// 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 BF16 = ck::bhalf_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 = BF16;
using BDataType = BF16;
using AccDataType = F32;
using CShuffleDataType = BF16;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
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
#include "run_grouped_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
View file @ 421b718d
...@@ -56,197 +56,6 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl ...@@ -56,197 +56,6 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
< 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>; < 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 // clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType, #include "run_grouped_gemm_example.inc"
BDataType,
EDataType,
AccDataType,
AElementOp,
BElementOp,
CDEElementOp>;
int main(int argc, char* argv[]) int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, 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;
}
example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp 0 → 100644
View file @ 421b718d
// 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 = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F32;
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, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>;
// clang-format on
#include "run_grouped_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp 0 → 100644
View file @ 421b718d
// 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 Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = ck::int4_t;
using BDataType = ck::int4_t;
using AccDataType = int32_t;
using CShuffleDataType = int32_t;
using DsDataType = ck::Tuple<>;
using EDataType = ck::int4_t;
using KernelADataType = int8_t;
using KernelBDataType = int8_t;
using KernelEDataType = int8_t;
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, //ALayout
BLayout, //BLayout
DsLayout, //DsLayout
ELayout, //ELayout
KernelADataType, //ADataType
KernelBDataType, //BDataType
AccDataType, //AccDataType
CShuffleDataType, //CShuffleDataType
DsDataType, //DsDataType
KernelEDataType, //EDataType
AElementOp, //AElementwiseOperation
BElementOp, //BElementwiseOperation
CDEElementOp, //CDEElementwiseOperation
GemmDefault, //GEMMSpecialization
1, // NumGemmKPrefetchStage
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
64, // KPerBlock
16, // AK1
16, // BK1
32, // MPerXdl
32, // NPerXdl
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransfer ThreadCluster Lengths_K0_M_K1
S<1, 0, 2>, // ABlockTransfer ThreadCluster ArrangeOrder
S<1, 0, 2>, // ABlockTransfer SrcAccessOrder
2, // ABlockTransfer SrcVectorDim
16, // ABlockTransfer SrcScalarPerVector
16, // ABlockTransfer DstScalarPerVector_K1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransfer ThreadCluster Lengths_K0_N_K1
S<1, 0, 2>, // BBlockTransfer ThreadCluster ArrangeOrder
S<1, 0, 2>, // BBlockTransfer SrcAccessOrder
2, // BBlockTransfer SrcVectorDim
16, // BBlockTransfer SrcScalarPerVector
16, // BBlockTransfer DstScalarPerVector_K1
1, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 64, 1, 4>, // CBlockTransferClusterLengths_MBlock_MWaveMPerXdl_NBlock_NWaveNPerXdl
16>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
#define BUILD_INT4_EXAMPLE
#include "run_grouped_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp 0 → 100644
View file @ 421b718d
// 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 Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = int8_t;
using BDataType = int8_t;
using AccDataType = int32_t;
using CShuffleDataType = int8_t;
using DsDataType = ck::Tuple<>;
using EDataType = int8_t;
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, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 1, 1, 1, S<1, 64, 1, 4>, 16>;
// clang-format on
#include "run_grouped_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
example/15_grouped_gemm/run_grouped_gemm_example.inc 0 → 100644
View file @ 421b718d
#pragma once
struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
static_assert(sizeof(ADataType) == sizeof(KernelADataType));
static_assert(sizeof(BDataType) == sizeof(KernelBDataType));
static_assert(sizeof(EDataType) == sizeof(KernelEDataType));
#endif
int group_count = problem_size.group_count;
// 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 = problem_size.Ms[i];
int N = problem_size.Ns[i];
int K = problem_size.Ks[i];
int stride_A = problem_size.stride_As[i];
int stride_B = problem_size.stride_Bs[i];
int stride_C = problem_size.stride_Cs[i];
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;
#ifdef BUILD_INT4_EXAMPLE
std::vector<Tensor<KernelEDataType>> c_device_tensors;
#else
std::vector<Tensor<EDataType>> c_device_tensors;
#endif
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{})));
#ifdef BUILD_INT4_EXAMPLE
c_device_tensors.push_back(Tensor<KernelEDataType>(f_host_tensor_descriptor(
gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
#else
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{})));
#endif
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(config.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()));
#ifdef BUILD_INT4_EXAMPLE
const Tensor<KernelADataType> a_converted(a_tensors[i]);
const Tensor<KernelBDataType> b_converted(b_tensors[i]);
a_tensors_device[i]->ToDevice(a_converted.mData.data());
b_tensors_device[i]->ToDevice(b_converted.mData.data());
#else
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
#endif
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");
}
invoker.Run(argument, StreamConfig{nullptr, false});
bool pass = true;
if(config.do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
EDataType,
AccDataType,
AElementOp,
BElementOp,
CDEElementOp>;
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);
#ifdef BUILD_INT4_EXAMPLE
const Tensor<EDataType> c_device_result_converted(c_device_tensors[i]);
pass &= ck::utils::check_err(c_device_result_converted.mData, c_host_tensors[i].mData);
#else
pass &= ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
#endif
}
}
if(config.time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.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;
}
return pass;
}
bool run_grouped_gemm_example(int argc, char* argv[])
{
ProblemSize problem_size;
ExecutionConfig config;
problem_size.group_count = 16;
for(int i = 0; i < problem_size.group_count; i++)
{
problem_size.Ms.push_back(256 + 256 * i);
problem_size.Ns.push_back(128 + 128 * i);
problem_size.Ks.push_back(128 + 64 * i);
problem_size.stride_As.push_back(problem_size.Ks[i]);
problem_size.stride_Bs.push_back(problem_size.Ks[i]);
problem_size.stride_Cs.push_back(problem_size.Ns[i]);
}
if(argc == 4)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
config.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);
}
return run_grouped_gemm(problem_size, config);
}
example/22_cgemm/CMakeLists.txt
View file @ 421b718d
...@@ -5,7 +5,13 @@ add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp) ...@@ -5,7 +5,13 @@ add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp) add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp)
add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp) add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16) add_dependencies(example_cgemm_xdl
add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16) example_cgemm_xdl_bf16
add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32) example_cgemm_xdl_fp16
add_dependencies(example_cgemm_xdl example_cgemm_xdl_int8) example_cgemm_xdl_fp32
example_cgemm_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)
endif()
example/22_cgemm/cgemm_xdl_bf16.cpp
View file @ 421b718d
...@@ -117,16 +117,16 @@ int main(int argc, char* argv[]) ...@@ -117,16 +117,16 @@ int main(int argc, char* argv[])
exit(0); exit(0);
} }
return run_cgemm_xdl<ADataType, return !run_cgemm_xdl<ADataType,
BDataType, BDataType,
CDataType, CDataType,
ALayout, ALayout,
BLayout, BLayout,
CLayout, CLayout,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough, PassThrough,
DeviceCGemmInstance, DeviceCGemmInstance,
ReferenceCGemmInstance>( ReferenceCGemmInstance>(
M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel); M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
} }
example/22_cgemm/cgemm_xdl_common.hpp
View file @ 421b718d
...@@ -21,6 +21,9 @@ using F32 = float; ...@@ -21,6 +21,9 @@ using F32 = float;
using BF16 = ck::bhalf_t; using BF16 = ck::bhalf_t;
using INT8 = std::int8_t; using INT8 = std::int8_t;
using INT32 = std::int32_t; using INT32 = std::int32_t;
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
using INT4 = ck::int4_t;
#endif
template <typename ADataType, template <typename ADataType,
typename BDataType, typename BDataType,
...@@ -32,17 +35,31 @@ template <typename ADataType, ...@@ -32,17 +35,31 @@ template <typename ADataType,
typename BElementwiseOperation, typename BElementwiseOperation,
typename CElementwiseOperation, typename CElementwiseOperation,
typename DeviceCGemmInstance, typename DeviceCGemmInstance,
typename ReferenceCGemmInstance> typename ReferenceCGemmInstance,
int run_cgemm_xdl(ck::index_t M, typename KernelADataType = ADataType,
ck::index_t N, typename KernelBDataType = BDataType,
ck::index_t K, typename KernelCDataType = CDataType>
ck::index_t StrideA, bool run_cgemm_xdl(ck::index_t M,
ck::index_t StrideB, ck::index_t N,
ck::index_t StrideC, ck::index_t K,
bool do_verification, ck::index_t StrideA,
int init_method, ck::index_t StrideB,
bool time_kernel) ck::index_t StrideC,
bool do_verification,
int init_method,
bool time_kernel)
{ {
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
static_assert(sizeof(ck::int4_t) == sizeof(int8_t),
"sizeof ck::int4_t and int8_t is different!");
static_assert(sizeof(ADataType) == sizeof(KernelADataType),
"sizeof ADataType and KernelADataType is different!");
static_assert(sizeof(BDataType) == sizeof(KernelBDataType),
"sizeof BDataType and KernelBDataType is different!");
static_assert(sizeof(CDataType) == sizeof(KernelCDataType),
"sizeof CDataType and KernelCDataType is different!");
#endif
auto f_host_tensor_descriptor = auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) { [](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) if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
...@@ -61,8 +78,10 @@ int run_cgemm_xdl(ck::index_t M, ...@@ -61,8 +78,10 @@ int run_cgemm_xdl(ck::index_t M,
Tensor<ADataType> a_m_k_imag(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); Tensor<ADataType> a_m_k_imag(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n_real(f_host_tensor_descriptor(K, N, StrideB, BLayout{})); Tensor<BDataType> b_k_n_real(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<BDataType> b_k_n_imag(f_host_tensor_descriptor(K, N, StrideB, BLayout{})); Tensor<BDataType> b_k_n_imag(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_real_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<KernelCDataType> c_m_n_real_device_result(
Tensor<CDataType> c_m_n_imag_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<KernelCDataType> c_m_n_imag_device_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k_real: " << a_m_k_real.mDesc << std::endl; std::cout << "a_m_k_real: " << a_m_k_real.mDesc << std::endl;
std::cout << "a_m_k_imag: " << a_m_k_imag.mDesc << std::endl; std::cout << "a_m_k_imag: " << a_m_k_imag.mDesc << std::endl;
...@@ -89,20 +108,41 @@ int run_cgemm_xdl(ck::index_t M, ...@@ -89,20 +108,41 @@ int run_cgemm_xdl(ck::index_t M,
auto cgemm = DeviceCGemmInstance{}; auto cgemm = DeviceCGemmInstance{};
DeviceMem a_m_k_real_device_buf(sizeof(ADataType) * a_m_k_real.mDesc.GetElementSpaceSize()); DeviceMem a_m_k_real_device_buf(sizeof(KernelADataType) *
DeviceMem a_m_k_imag_device_buf(sizeof(ADataType) * a_m_k_imag.mDesc.GetElementSpaceSize()); a_m_k_real.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_real_device_buf(sizeof(BDataType) * b_k_n_real.mDesc.GetElementSpaceSize()); DeviceMem a_m_k_imag_device_buf(sizeof(KernelADataType) *
DeviceMem b_k_n_imag_device_buf(sizeof(BDataType) * b_k_n_imag.mDesc.GetElementSpaceSize()); a_m_k_imag.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_real_device_buf(sizeof(CDataType) * DeviceMem b_k_n_real_device_buf(sizeof(KernelBDataType) *
b_k_n_real.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_imag_device_buf(sizeof(KernelBDataType) *
b_k_n_imag.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_real_device_buf(sizeof(KernelCDataType) *
c_m_n_real_device_result.mDesc.GetElementSpaceSize()); c_m_n_real_device_result.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_imag_device_buf(sizeof(CDataType) * DeviceMem c_m_n_imag_device_buf(sizeof(KernelCDataType) *
c_m_n_imag_device_result.mDesc.GetElementSpaceSize()); c_m_n_imag_device_result.mDesc.GetElementSpaceSize());
DeviceMem workspace_device_buf(cgemm.GetWorkspaceSize(M, N, K, StrideA, StrideB, StrideC)); DeviceMem workspace_device_buf(cgemm.GetWorkspaceSize(M, N, K, StrideA, StrideB, StrideC));
a_m_k_real_device_buf.ToDevice(a_m_k_real.mData.data()); #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
a_m_k_imag_device_buf.ToDevice(a_m_k_imag.mData.data()); if constexpr(std::is_same_v<ADataType, ck::int4_t>)
b_k_n_real_device_buf.ToDevice(b_k_n_real.mData.data()); {
b_k_n_imag_device_buf.ToDevice(b_k_n_imag.mData.data()); Tensor<KernelADataType> a_m_k_real_converted(a_m_k_real);
Tensor<KernelADataType> a_m_k_imag_converted(a_m_k_imag);
Tensor<KernelBDataType> b_k_n_real_converted(b_k_n_real);
Tensor<KernelBDataType> b_k_n_imag_converted(b_k_n_imag);
a_m_k_real_device_buf.ToDevice(a_m_k_real_converted.mData.data());
a_m_k_imag_device_buf.ToDevice(a_m_k_imag_converted.mData.data());
b_k_n_real_device_buf.ToDevice(b_k_n_real_converted.mData.data());
b_k_n_imag_device_buf.ToDevice(b_k_n_imag_converted.mData.data());
}
else
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
{
a_m_k_real_device_buf.ToDevice(a_m_k_real.mData.data());
a_m_k_imag_device_buf.ToDevice(a_m_k_imag.mData.data());
b_k_n_real_device_buf.ToDevice(b_k_n_real.mData.data());
b_k_n_imag_device_buf.ToDevice(b_k_n_imag.mData.data());
}
auto a_element_op = AElementwiseOperation{}; auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{}; auto b_element_op = BElementwiseOperation{};
...@@ -111,13 +151,13 @@ int run_cgemm_xdl(ck::index_t M, ...@@ -111,13 +151,13 @@ int run_cgemm_xdl(ck::index_t M,
// do GEMM // do GEMM
auto invoker = cgemm.MakeInvoker(); auto invoker = cgemm.MakeInvoker();
auto argument = auto argument =
cgemm.MakeArgument(static_cast<ADataType*>(a_m_k_real_device_buf.GetDeviceBuffer()), cgemm.MakeArgument(static_cast<KernelADataType*>(a_m_k_real_device_buf.GetDeviceBuffer()),
static_cast<ADataType*>(a_m_k_imag_device_buf.GetDeviceBuffer()), static_cast<KernelADataType*>(a_m_k_imag_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_real_device_buf.GetDeviceBuffer()), static_cast<KernelBDataType*>(b_k_n_real_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_imag_device_buf.GetDeviceBuffer()), static_cast<KernelBDataType*>(b_k_n_imag_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_real_device_buf.GetDeviceBuffer()), static_cast<KernelCDataType*>(c_m_n_real_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_imag_device_buf.GetDeviceBuffer()), static_cast<KernelCDataType*>(c_m_n_imag_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(workspace_device_buf.GetDeviceBuffer()), static_cast<KernelCDataType*>(workspace_device_buf.GetDeviceBuffer()),
M, M,
N, N,
K, K,
...@@ -142,16 +182,12 @@ int run_cgemm_xdl(ck::index_t M, ...@@ -142,16 +182,12 @@ int run_cgemm_xdl(ck::index_t M,
std::size_t(2) * std::size_t(2) *
(sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N); (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time; float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< cgemm.GetTypeString() << std::endl; << cgemm.GetTypeString() << std::endl;
c_m_n_real_device_buf.FromDevice(c_m_n_real_device_result.mData.data());
c_m_n_imag_device_buf.FromDevice(c_m_n_imag_device_result.mData.data());
if(do_verification) if(do_verification)
{ {
Tensor<CDataType> c_m_n_real_host_result( Tensor<CDataType> c_m_n_real_host_result(
...@@ -159,9 +195,8 @@ int run_cgemm_xdl(ck::index_t M, ...@@ -159,9 +195,8 @@ int run_cgemm_xdl(ck::index_t M,
Tensor<CDataType> c_m_n_imag_host_result( Tensor<CDataType> c_m_n_imag_host_result(
f_host_tensor_descriptor(M, N, StrideC, CLayout{})); f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
auto ref_cgemm = ReferenceCGemmInstance{}; auto ref_cgemm = ReferenceCGemmInstance{};
auto ref_invoker = ref_cgemm.MakeInvoker(); auto ref_invoker = ref_cgemm.MakeInvoker();
auto ref_argument = ref_cgemm.MakeArgument(a_m_k_real, auto ref_argument = ref_cgemm.MakeArgument(a_m_k_real,
a_m_k_imag, a_m_k_imag,
b_k_n_real, b_k_n_real,
...@@ -174,19 +209,45 @@ int run_cgemm_xdl(ck::index_t M, ...@@ -174,19 +209,45 @@ int run_cgemm_xdl(ck::index_t M,
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
c_m_n_real_device_buf.FromDevice(c_m_n_real_device_result.mData.data());
c_m_n_imag_device_buf.FromDevice(c_m_n_imag_device_result.mData.data());
bool result = true; bool result = true;
result = ck::utils::check_err(c_m_n_real_device_result.mData, #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
c_m_n_real_host_result.mData, if constexpr(std::is_same_v<ADataType, ck::int4_t>)
"Verification error: incorrect results in real part!", {
1e-2f, const Tensor<CDataType> c_m_n_real_device_result_converted(c_m_n_real_device_result);
1e-1f); const Tensor<CDataType> c_m_n_imag_device_result_converted(c_m_n_imag_device_result);
result = result &&
ck::utils::check_err(c_m_n_imag_device_result.mData, result = ck::utils::check_err(c_m_n_real_device_result_converted.mData,
c_m_n_imag_host_result.mData, c_m_n_real_host_result.mData,
"Verification error: incorrect results in imaginary part!", "Verification error: incorrect results in real part!",
1e-2f, 1e-2f,
1e-1f); 1e-1f);
return result ? 0 : 1; result = result && ck::utils::check_err(
c_m_n_imag_device_result_converted.mData,
c_m_n_imag_host_result.mData,
"Verification error: incorrect results in imaginary part!",
1e-2f,
1e-1f);
}
else
#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
{
result = ck::utils::check_err(c_m_n_real_device_result.mData,
c_m_n_real_host_result.mData,
"Verification error: incorrect results in real part!",
1e-2f,
1e-1f);
result = result && ck::utils::check_err(
c_m_n_imag_device_result.mData,
c_m_n_imag_host_result.mData,
"Verification error: incorrect results in imaginary part!",
1e-2f,
1e-1f);
}
return result;
} }
return 0; return true;
} }
example/22_cgemm/cgemm_xdl_fp16.cpp
View file @ 421b718d
...@@ -116,16 +116,16 @@ int main(int argc, char* argv[]) ...@@ -116,16 +116,16 @@ int main(int argc, char* argv[])
exit(0); exit(0);
} }
return run_cgemm_xdl<ADataType, return !run_cgemm_xdl<ADataType,
BDataType, BDataType,
CDataType, CDataType,
ALayout, ALayout,
BLayout, BLayout,
CLayout, CLayout,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough, PassThrough,
DeviceCGemmInstance, DeviceCGemmInstance,
ReferenceCGemmInstance>( ReferenceCGemmInstance>(
M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel); M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
} }
example/22_cgemm/cgemm_xdl_fp32.cpp
View file @ 421b718d
...@@ -117,16 +117,16 @@ int main(int argc, char* argv[]) ...@@ -117,16 +117,16 @@ int main(int argc, char* argv[])
exit(0); exit(0);
} }
return run_cgemm_xdl<ADataType, return !run_cgemm_xdl<ADataType,
BDataType, BDataType,
CDataType, CDataType,
ALayout, ALayout,
BLayout, BLayout,
CLayout, CLayout,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough, PassThrough,
DeviceCGemmInstance, DeviceCGemmInstance,
ReferenceCGemmInstance>( ReferenceCGemmInstance>(
M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel); M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
} }
example/22_cgemm/cgemm_xdl_int4.cpp 0 → 100644
View file @ 421b718d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include "cgemm_xdl_common.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_cgemm_4gemm_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
using ADataType = INT4;
using BDataType = INT4;
using CDataType = INT4;
using AccDataType = INT32;
using CShuffleDataType = INT32;
using KernelADataType = INT8;
using KernelBDataType = INT8;
using KernelCDataType = INT8;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
using ReferenceCGemmInstance = ck::tensor_operation::host::
ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
// clang-format off
using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
<ALayout, // typename ALayout
BLayout, // typename BLayout
CLayout, // typename CLayout
KernelADataType, // typename ADataType
KernelBDataType, // typename BDataType
KernelCDataType, // typename CDataType
AccDataType, // typename GemmAccDataType
CShuffleDataType, // typename CShuffleDataType
PassThrough, // typename AElementwiseOperation
PassThrough, // typename BElementwiseOperation
PassThrough, // typename CElementwiseOperation
GemmDefault, // GemmSpecialization GemmSpec
1, // index_t NumGemmKPrefetchStage
256, // index_t BlockSize
256, // index_t MPerBlock
128, // index_t NPerBlock
64, // index_t KPerBlock
16, // index_t AK1
16, // index_t BK1
32, // index_t MPerXDL
32, // index_t NPerXDL
4, // index_t MXdlPerWave
2, // index_t NXdlPerWave
S<4, 64, 1>, // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // typename ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // typename ABlockTransferSrcAccessOrder
2, // index_t ABlockTransferSrcVectorDim
16, // index_t ABlockTransferSrcScalarPerVector
16, // index_t ABlockTransferDstScalarPerVector_AK1
1, // index_t ABlockLdsExtraM
S<4, 64, 1>, // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // typename BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // typename BBlockTransferSrcAccessOrder
2, // index_t BBlockTransferSrcVectorDim
8, // index_t BBlockTransferSrcScalarPerVector
8, // index_t BBlockTransferDstScalarPerVector_BK1
1, // index_t BBlockLdsExtraN
1, // index_t CShuffleMXdlPerWavePerShuffle
1, // index_t CShuffleNXdlPerWavePerShuffle
S<1, 64, 1, 4>, // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
16>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
// clang-format on
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
// CGEMM shape
ck::index_t M = 1024;
ck::index_t N = 1152;
ck::index_t K = 512;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideC = N;
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
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
<< std::endl;
exit(EXIT_SUCCESS);
}
return !run_cgemm_xdl<ADataType,
BDataType,
CDataType,
ALayout,
BLayout,
CLayout,
PassThrough,
PassThrough,
PassThrough,
DeviceCGemmInstance,
ReferenceCGemmInstance,
KernelADataType,
KernelBDataType,
KernelCDataType>(
M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment