mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-20 21:09:08 +00:00
gemm, conv perchannel quantization (#503)
* Use gemm_multiple_D instead
* Add gemm bias relu quantization example
* Add pure gemm quantization example
* Add quantization of perchannel conv + bias + relu example
* Refine the code
* Rename multiplier to requant_scale
* Rename the folder
* Remove redundant comment
* Rename the file. Prepare to add perchannel
* Add conv perchannel instance
* Move to quantization folder
* Add conv perchannel client example
* Apply Rangify constructor of HostTensorDescriptor & Tensor<>
* Fix merge error
[ROCm/composable_kernel commit: ad541ad6b9]
This commit is contained in:
rocking5566
2
example/14_gemm_quantization/CMakeLists.txt
Normal file
2
example/14_gemm_quantization/CMakeLists.txt
Normal file
@@ -0,0 +1,2 @@
|
||||
add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp)
|
||||
add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp)
|
||||
@@ -0,0 +1,235 @@
|
||||
// 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/impl/device_gemm_multiple_d_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 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 ActivationOp = ck::tensor_operation::element_wise::Relu;
|
||||
using CDEElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
using ADataType = I8;
|
||||
using BDataType = I8;
|
||||
using AccDataType = I32;
|
||||
using CShuffleDataType = I32;
|
||||
using BiasDataType = I32;
|
||||
using DsDataType = ck::Tuple<BiasDataType>;
|
||||
using EDataType = I8;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Col;
|
||||
using BiasLayout = Row;
|
||||
using DsLayout = ck::Tuple<BiasLayout>;
|
||||
using ELayout = Row;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
// clang-format off
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<
|
||||
ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
PassThrough, // AElementwiseOperation,
|
||||
PassThrough, // BElementwiseOperation,
|
||||
CDEElementOp, // CDEElementwiseOperation,
|
||||
GemmDefault, // GemmSpecialization GemmSpec,
|
||||
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>, // ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
|
||||
S<1, 0, 2>, // ABlockTransferSrcAccessOrder,
|
||||
2, // index_t ABlockTransferSrcVectorDim,
|
||||
16, // index_t ABlockTransferSrcScalarPerVector,
|
||||
16, // index_t ABlockTransferDstScalarPerVector_AK1,
|
||||
1, // bool 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, // bool BBlockLdsExtraN,
|
||||
1, // index_t CShuffleMXdlPerWavePerShuffle,
|
||||
1, // index_t CShuffleNXdlPerWavePerShuffle,
|
||||
S<1, 64, 1, 4>, // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
|
||||
8>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
|
||||
// clang-format on
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
AccDataType,
|
||||
PassThrough,
|
||||
PassThrough,
|
||||
PassThrough>;
|
||||
|
||||
int main()
|
||||
{
|
||||
bool do_verification = true;
|
||||
bool time_kernel = false;
|
||||
|
||||
// GEMM shape
|
||||
ck::index_t M = 1024;
|
||||
ck::index_t N = 1024;
|
||||
ck::index_t K = 1024;
|
||||
|
||||
ck::index_t StrideA = 1024;
|
||||
ck::index_t StrideB = 1024;
|
||||
ck::index_t StrideBias = 0;
|
||||
ck::index_t StrideE = 1024;
|
||||
|
||||
float requant_scale = 0.03;
|
||||
|
||||
auto f_host_tensor_descriptor2d =
|
||||
[](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(std::vector<std::size_t>({row, col}),
|
||||
std::vector<std::size_t>({stride, 1_uz}));
|
||||
}
|
||||
else
|
||||
{
|
||||
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
|
||||
std::vector<std::size_t>({1_uz, stride}));
|
||||
}
|
||||
};
|
||||
|
||||
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
|
||||
return HostTensorDescriptor(std::vector<std::size_t>({len}),
|
||||
std::vector<std::size_t>({stride}));
|
||||
};
|
||||
|
||||
Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
|
||||
Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
|
||||
Tensor<BiasDataType> bias_n(f_host_tensor_descriptor1d(N, 1));
|
||||
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
|
||||
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
|
||||
|
||||
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
|
||||
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
|
||||
std::cout << "bias_n: " << bias_n.mDesc << std::endl;
|
||||
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
|
||||
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-128, 127});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-128, 127});
|
||||
bias_n.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-128, 127});
|
||||
|
||||
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
a_device_buf.ToDevice(a_m_k.mData.data());
|
||||
b_device_buf.ToDevice(b_k_n.mData.data());
|
||||
bias_device_buf.ToDevice(bias_n.mData.data());
|
||||
|
||||
auto a_element_op = PassThrough{};
|
||||
auto b_element_op = PassThrough{};
|
||||
auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
|
||||
|
||||
// do GEMM
|
||||
auto gemm = DeviceGemmInstance{};
|
||||
auto invoker = gemm.MakeInvoker();
|
||||
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
|
||||
b_device_buf.GetDeviceBuffer(),
|
||||
{bias_device_buf.GetDeviceBuffer()},
|
||||
e_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
{StrideBias},
|
||||
StrideE,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
if(!gemm.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
}
|
||||
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
std::size_t num_btype =
|
||||
sizeof(ADataType) * M * K + sizeof(BDataType) * K * 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, "
|
||||
<< gemm.GetTypeString() << std::endl;
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
|
||||
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument =
|
||||
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
for(int m = 0; m < M; ++m)
|
||||
{
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), bias_n(n));
|
||||
}
|
||||
}
|
||||
|
||||
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
207
example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
Normal file
207
example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
Normal file
@@ -0,0 +1,207 @@
|
||||
// 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/impl/device_gemm_multiple_d_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 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 ActivationOp = PassThrough;
|
||||
using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
using ADataType = I8;
|
||||
using BDataType = I8;
|
||||
using AccDataType = I32;
|
||||
using CShuffleDataType = I32;
|
||||
using DsDataType = ck::Tuple<>;
|
||||
using EDataType = I8;
|
||||
|
||||
using ALayout = Row;
|
||||
using BLayout = Col;
|
||||
using DsLayout = ck::Tuple<>;
|
||||
using ELayout = Row;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
// clang-format off
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<
|
||||
ALayout,
|
||||
BLayout,
|
||||
DsLayout,
|
||||
ELayout,
|
||||
ADataType,
|
||||
BDataType,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
DsDataType,
|
||||
EDataType,
|
||||
PassThrough, // AElementwiseOperation,
|
||||
PassThrough, // BElementwiseOperation,
|
||||
CDEElementOp, // CDEElementwiseOperation,
|
||||
GemmDefault, // GemmSpecialization GemmSpec,
|
||||
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>, // ABlockTransferThreadClusterLengths_AK0_M_AK1,
|
||||
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
|
||||
S<1, 0, 2>, // ABlockTransferSrcAccessOrder,
|
||||
2, // index_t ABlockTransferSrcVectorDim,
|
||||
16, // index_t ABlockTransferSrcScalarPerVector,
|
||||
16, // index_t ABlockTransferDstScalarPerVector_AK1,
|
||||
1, // bool 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, // bool 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
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::
|
||||
ReferenceGemm<ADataType, BDataType, EDataType, float, PassThrough, PassThrough, CDEElementOp>;
|
||||
|
||||
int main()
|
||||
{
|
||||
bool do_verification = true;
|
||||
bool time_kernel = false;
|
||||
|
||||
// GEMM shape
|
||||
ck::index_t M = 1024;
|
||||
ck::index_t N = 1024;
|
||||
ck::index_t K = 1024;
|
||||
|
||||
ck::index_t StrideA = 1024;
|
||||
ck::index_t StrideB = 1024;
|
||||
ck::index_t StrideE = 1024;
|
||||
|
||||
float requant_scale = 0.03;
|
||||
|
||||
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(std::vector<std::size_t>({row, col}),
|
||||
std::vector<std::size_t>({stride, 1_uz}));
|
||||
}
|
||||
else
|
||||
{
|
||||
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
|
||||
std::vector<std::size_t>({1_uz, stride}));
|
||||
}
|
||||
};
|
||||
|
||||
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
|
||||
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
|
||||
|
||||
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
|
||||
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
|
||||
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
|
||||
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-128, 127});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-128, 127});
|
||||
|
||||
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
a_device_buf.ToDevice(a_m_k.mData.data());
|
||||
b_device_buf.ToDevice(b_k_n.mData.data());
|
||||
|
||||
auto a_element_op = PassThrough{};
|
||||
auto b_element_op = PassThrough{};
|
||||
auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
|
||||
|
||||
// do GEMM
|
||||
auto gemm = DeviceGemmInstance{};
|
||||
auto invoker = gemm.MakeInvoker();
|
||||
auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
|
||||
b_device_buf.GetDeviceBuffer(),
|
||||
{},
|
||||
e_device_buf.GetDeviceBuffer(),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
{},
|
||||
StrideE,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
cde_element_op);
|
||||
|
||||
if(!gemm.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
}
|
||||
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
std::size_t num_btype =
|
||||
sizeof(ADataType) * M * K + sizeof(BDataType) * K * 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, "
|
||||
<< gemm.GetTypeString() << std::endl;
|
||||
|
||||
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(
|
||||
a_m_k, b_k_n, e_m_n_host_result, a_element_op, b_element_op, cde_element_op);
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
@@ -1 +0,0 @@
|
||||
add_example_executable(example_gemm_xdl_relu_quantization_int8 gemm_xdl_relu_quantization_int8.cpp)
|
||||
@@ -1,233 +0,0 @@
|
||||
// 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/impl/device_gemm_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 PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using ActivationOp = ck::tensor_operation::element_wise::Relu;
|
||||
using CElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
|
||||
|
||||
using ADataType = int8_t;
|
||||
using BDataType = int8_t;
|
||||
using CDataType = int8_t;
|
||||
using AccDataType = int32_t;
|
||||
using CShuffleDataType = float;
|
||||
|
||||
using ALayout = ck::tensor_layout::gemm::RowMajor;
|
||||
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
|
||||
using CLayout = ck::tensor_layout::gemm::RowMajor;
|
||||
|
||||
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
|
||||
|
||||
// clang-format off
|
||||
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle<
|
||||
ALayout, // typename ALayout,
|
||||
BLayout, // typename BLayout,
|
||||
CLayout, // typename CLayout,
|
||||
ADataType, // typename ADataType,
|
||||
BDataType, // typename BDataType,
|
||||
CDataType, // typename CDataType,
|
||||
AccDataType, // typename GemmAccDataType,
|
||||
CShuffleDataType, // typename CShuffleDataType,
|
||||
PassThrough, // typename AElementwiseOperation,
|
||||
PassThrough, // typename BElementwiseOperation,
|
||||
CElementOp, // 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, // bool 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, // bool 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
|
||||
|
||||
using ReferenceGemmInstance = ck::tensor_operation::host::
|
||||
ReferenceGemm<ADataType, BDataType, CDataType, float, PassThrough, PassThrough, CElementOp>;
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool do_verification = true;
|
||||
int init_method = 1;
|
||||
bool time_kernel = false;
|
||||
|
||||
// GEMM shape
|
||||
ck::index_t M = 3840;
|
||||
ck::index_t N = 4096;
|
||||
ck::index_t K = 4096;
|
||||
|
||||
ck::index_t StrideA = 4096;
|
||||
ck::index_t StrideB = 4096;
|
||||
ck::index_t StrideC = 4096;
|
||||
|
||||
float quant_multiplier = 0.03;
|
||||
|
||||
if(argc == 4)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
}
|
||||
else if(argc == 10)
|
||||
{
|
||||
do_verification = std::stoi(argv[1]);
|
||||
init_method = std::stoi(argv[2]);
|
||||
time_kernel = std::stoi(argv[3]);
|
||||
|
||||
M = std::stoi(argv[4]);
|
||||
N = std::stoi(argv[5]);
|
||||
K = std::stoi(argv[6]);
|
||||
|
||||
StrideA = std::stoi(argv[7]);
|
||||
StrideB = std::stoi(argv[8]);
|
||||
StrideC = std::stoi(argv[9]);
|
||||
}
|
||||
else
|
||||
{
|
||||
printf("arg1: verification (0=no, 1=yes)\n");
|
||||
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
|
||||
printf("arg3: time kernel (0=n0, 1=yes)\n");
|
||||
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
|
||||
exit(0);
|
||||
}
|
||||
|
||||
auto f_host_tensor_descriptor =
|
||||
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
|
||||
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<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
|
||||
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
|
||||
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
|
||||
|
||||
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
|
||||
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
|
||||
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
|
||||
break;
|
||||
default:
|
||||
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
|
||||
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
|
||||
}
|
||||
|
||||
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
|
||||
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
|
||||
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
|
||||
|
||||
a_m_k_device_buf.ToDevice(a_m_k.mData.data());
|
||||
b_k_n_device_buf.ToDevice(b_k_n.mData.data());
|
||||
|
||||
auto a_element_op = PassThrough{};
|
||||
auto b_element_op = PassThrough{};
|
||||
auto c_element_op = CElementOp{quant_multiplier, ActivationOp{}};
|
||||
|
||||
// do GEMM
|
||||
auto gemm = DeviceGemmInstance{};
|
||||
auto invoker = gemm.MakeInvoker();
|
||||
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
|
||||
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
|
||||
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
|
||||
M,
|
||||
N,
|
||||
K,
|
||||
StrideA,
|
||||
StrideB,
|
||||
StrideC,
|
||||
a_element_op,
|
||||
b_element_op,
|
||||
c_element_op);
|
||||
|
||||
if(!gemm.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_gemm with the specified compilation parameters does "
|
||||
"not support this GEMM problem");
|
||||
}
|
||||
|
||||
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = std::size_t(2) * M * N * K;
|
||||
std::size_t num_btype =
|
||||
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
|
||||
|
||||
float gb_per_sec = num_btype / 1.E6 / ave_time;
|
||||
|
||||
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
|
||||
<< gemm.GetTypeString() << std::endl;
|
||||
|
||||
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
auto ref_gemm = ReferenceGemmInstance{};
|
||||
auto ref_invoker = ref_gemm.MakeInvoker();
|
||||
|
||||
auto ref_argument = ref_gemm.MakeArgument(
|
||||
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
return ck::utils::check_err(c_m_n_device_result, c_m_n_host_result) ? 0 : 1;
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
@@ -1,2 +1,3 @@
|
||||
add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp)
|
||||
add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp)
|
||||
add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp)
|
||||
@@ -0,0 +1,342 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
|
||||
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
|
||||
|
||||
#include "ck/library/utility/algorithm.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"
|
||||
#include "ck/library/utility/convolution_parameter.hpp"
|
||||
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
|
||||
|
||||
using InDataType = int8_t;
|
||||
using WeiDataType = int8_t;
|
||||
using BiasDataType = int32_t;
|
||||
using RequantScaleDataType = float;
|
||||
using AccDataType = int32_t;
|
||||
using CShuffleDataType = int32_t;
|
||||
using OutDataType = int8_t;
|
||||
|
||||
template <ck::index_t... Is>
|
||||
using S = ck::Sequence<Is...>;
|
||||
|
||||
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
|
||||
using InElementOp = PassThrough;
|
||||
using WeiElementOp = PassThrough;
|
||||
using ActivationOp = ck::tensor_operation::element_wise::Relu;
|
||||
using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
|
||||
|
||||
static constexpr auto ConvSpec =
|
||||
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
|
||||
|
||||
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
|
||||
|
||||
template <ck::index_t NDimSpatial,
|
||||
typename InLayout,
|
||||
typename WeiLayout,
|
||||
typename BiasLayout,
|
||||
typename RequantScaleLayout,
|
||||
typename OutLayout>
|
||||
using DeviceGroupedConvNDFwdInstance =
|
||||
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
|
||||
NDimSpatial,
|
||||
InLayout,
|
||||
WeiLayout,
|
||||
ck::Tuple<BiasLayout, RequantScaleLayout>,
|
||||
OutLayout,
|
||||
InDataType,
|
||||
WeiDataType,
|
||||
AccDataType,
|
||||
CShuffleDataType,
|
||||
ck::Tuple<BiasDataType, RequantScaleDataType>,
|
||||
OutDataType,
|
||||
InElementOp,
|
||||
WeiElementOp,
|
||||
OutElementOp,
|
||||
ConvSpec, // ConvForwardSpecialization
|
||||
GemmSpec, // GemmSpecialization
|
||||
1, //
|
||||
256, // BlockSize
|
||||
128, // MPerBlock
|
||||
256, // NPerBlock
|
||||
64, // KPerBlock
|
||||
16, // AK1
|
||||
16, // BK1
|
||||
32, // MPerXdl
|
||||
32, // NPerXdl
|
||||
2, // MXdlPerWave
|
||||
4, // NXdlPerWave
|
||||
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
|
||||
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
|
||||
2, // ABlockTransferSrcVectorDim
|
||||
16, // ABlockTransferSrcScalarPerVector
|
||||
16, // ABlockTransferDstScalarPerVector_AK1
|
||||
1, // ABlockLdsExtraM
|
||||
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
|
||||
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
|
||||
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
|
||||
2, // BBlockTransferSrcVectorDim
|
||||
16, // BBlockTransferSrcScalarPerVector
|
||||
16, // BBlockTransferDstScalarPerVector_BK1
|
||||
1, // BBlockLdsExtraN
|
||||
1,
|
||||
1,
|
||||
S<1, 64, 1, 4>,
|
||||
8>;
|
||||
|
||||
template <ck::index_t NDimSpatial,
|
||||
typename InDataType,
|
||||
typename WeiDataType,
|
||||
typename OutDataType,
|
||||
typename InElementOp,
|
||||
typename WeiElementOp,
|
||||
typename OutElementOp,
|
||||
typename DeviceConvNDFwdInstance>
|
||||
bool run_grouped_conv_fwd(bool do_verification,
|
||||
bool time_kernel,
|
||||
const ck::utils::conv::ConvParam& conv_param,
|
||||
const HostTensorDescriptor& in_g_n_c_wis_desc,
|
||||
const HostTensorDescriptor& wei_g_k_c_xs_desc,
|
||||
const HostTensorDescriptor& bias_g_k_desc,
|
||||
const HostTensorDescriptor& requant_scale_g_k_desc,
|
||||
const HostTensorDescriptor& out_g_n_k_wos_desc,
|
||||
const InElementOp& in_element_op,
|
||||
const WeiElementOp& wei_element_op,
|
||||
const OutElementOp& out_element_op)
|
||||
{
|
||||
Tensor<InDataType> in(in_g_n_c_wis_desc);
|
||||
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
|
||||
Tensor<BiasDataType> bias(bias_g_k_desc);
|
||||
Tensor<RequantScaleDataType> requant_scale(requant_scale_g_k_desc);
|
||||
Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
|
||||
Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
|
||||
|
||||
std::cout << "in: " << in.mDesc << std::endl;
|
||||
std::cout << "wei: " << wei.mDesc << std::endl;
|
||||
std::cout << "bias: " << bias.mDesc << std::endl;
|
||||
std::cout << "requant_scale: " << requant_scale.mDesc << std::endl;
|
||||
std::cout << "out: " << out_host.mDesc << std::endl;
|
||||
|
||||
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-128, 127});
|
||||
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-128, 127});
|
||||
bias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-128, 127});
|
||||
requant_scale.GenerateTensorValue(GeneratorTensor_2<RequantScaleDataType>{0, 1});
|
||||
|
||||
DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
|
||||
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
|
||||
DeviceMem bias_device_buf(sizeof(BiasDataType) * bias.mDesc.GetElementSpaceSize());
|
||||
DeviceMem requant_scale_device_buf(sizeof(RequantScaleDataType) *
|
||||
requant_scale.mDesc.GetElementSpaceSize());
|
||||
DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
|
||||
|
||||
in_device_buf.ToDevice(in.mData.data());
|
||||
wei_device_buf.ToDevice(wei.mData.data());
|
||||
bias_device_buf.ToDevice(bias.mData.data());
|
||||
requant_scale_device_buf.ToDevice(requant_scale.mData.data());
|
||||
|
||||
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
|
||||
std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
|
||||
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
|
||||
std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
|
||||
std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
|
||||
std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
|
||||
std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_lengths{};
|
||||
std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_strides{};
|
||||
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
|
||||
std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
|
||||
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
|
||||
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
|
||||
std::array<ck::index_t, NDimSpatial> input_left_pads{};
|
||||
std::array<ck::index_t, NDimSpatial> input_right_pads{};
|
||||
|
||||
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
|
||||
|
||||
copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
|
||||
copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
|
||||
copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
|
||||
copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
|
||||
copy(bias_g_k_desc.GetLengths(), d0_g_n_k_wos_lengths);
|
||||
copy(bias_g_k_desc.GetStrides(), d0_g_n_k_wos_strides);
|
||||
copy(requant_scale_g_k_desc.GetLengths(), d1_g_n_k_wos_lengths);
|
||||
copy(requant_scale_g_k_desc.GetStrides(), d1_g_n_k_wos_strides);
|
||||
copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
|
||||
copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
|
||||
copy(conv_param.conv_filter_strides_, conv_filter_strides);
|
||||
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
|
||||
copy(conv_param.input_left_pads_, input_left_pads);
|
||||
copy(conv_param.input_right_pads_, input_right_pads);
|
||||
|
||||
// do Conv
|
||||
auto conv = DeviceConvNDFwdInstance{};
|
||||
auto invoker = conv.MakeInvoker();
|
||||
auto argument = conv.MakeArgument(
|
||||
in_device_buf.GetDeviceBuffer(),
|
||||
wei_device_buf.GetDeviceBuffer(),
|
||||
{bias_device_buf.GetDeviceBuffer(), requant_scale_device_buf.GetDeviceBuffer()},
|
||||
out_device_buf.GetDeviceBuffer(),
|
||||
a_g_n_c_wis_lengths,
|
||||
a_g_n_c_wis_strides,
|
||||
b_g_k_c_xs_lengths,
|
||||
b_g_k_c_xs_strides,
|
||||
{d0_g_n_k_wos_lengths, d1_g_n_k_wos_lengths},
|
||||
{d0_g_n_k_wos_strides, d1_g_n_k_wos_strides},
|
||||
e_g_n_k_wos_lengths,
|
||||
e_g_n_k_wos_strides,
|
||||
conv_filter_strides,
|
||||
conv_filter_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads,
|
||||
in_element_op,
|
||||
wei_element_op,
|
||||
out_element_op);
|
||||
|
||||
if(!conv.IsSupportedArgument(argument))
|
||||
{
|
||||
throw std::runtime_error(
|
||||
"wrong! device_conv with the specified compilation parameters does "
|
||||
"not support this Conv problem");
|
||||
}
|
||||
|
||||
float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t flop = conv_param.GetFlops();
|
||||
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
|
||||
|
||||
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
|
||||
float gb_per_sec = num_btype / 1.E6 / avg_time;
|
||||
std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
|
||||
<< conv.GetTypeString() << std::endl;
|
||||
|
||||
bool pass = true;
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
Tensor<CShuffleDataType> c_host(out_g_n_k_wos_desc);
|
||||
|
||||
auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
|
||||
InDataType,
|
||||
WeiDataType,
|
||||
CShuffleDataType,
|
||||
InElementOp,
|
||||
WeiElementOp,
|
||||
PassThrough>();
|
||||
|
||||
auto ref_invoker = ref_conv.MakeInvoker();
|
||||
auto ref_argument = ref_conv.MakeArgument(in,
|
||||
wei,
|
||||
c_host,
|
||||
conv_param.conv_filter_strides_,
|
||||
conv_param.conv_filter_dilations_,
|
||||
conv_param.input_left_pads_,
|
||||
conv_param.input_right_pads_,
|
||||
in_element_op,
|
||||
wei_element_op,
|
||||
PassThrough{});
|
||||
|
||||
ref_invoker.Run(ref_argument);
|
||||
|
||||
// TODO: implement elementwise operation for host
|
||||
out_host.ForEach([&](auto&, auto idx) {
|
||||
out_element_op(out_host(idx), c_host(idx), bias(idx), requant_scale(idx));
|
||||
});
|
||||
|
||||
out_device_buf.FromDevice(out_device.mData.data());
|
||||
|
||||
pass &=
|
||||
ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
|
||||
}
|
||||
|
||||
return (pass ? 0 : 1);
|
||||
}
|
||||
|
||||
int main()
|
||||
{
|
||||
bool do_verification = true;
|
||||
bool time_kernel = true;
|
||||
const ck::index_t ndim_spatial = 2;
|
||||
|
||||
ck::utils::conv::ConvParam conv_param{
|
||||
ndim_spatial, // n_dim
|
||||
1, // group
|
||||
4, // batch
|
||||
64, // output channels
|
||||
32, // input chanels
|
||||
{3, 3}, // weight HW
|
||||
{71, 71}, // x HW
|
||||
{2, 2}, // strides
|
||||
{1, 1}, // dilations
|
||||
{1, 1}, // left_pads
|
||||
{1, 1} // right_pads
|
||||
};
|
||||
|
||||
const auto in_element_op = InElementOp{};
|
||||
const auto wei_element_op = WeiElementOp{};
|
||||
const auto out_element_op = OutElementOp{ActivationOp{}};
|
||||
|
||||
using InLayout = ck::tensor_layout::convolution::GNHWC;
|
||||
using WeiLayout = ck::tensor_layout::convolution::GKYXC;
|
||||
using BiasLayout = ck::tensor_layout::convolution::G_K;
|
||||
using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
|
||||
using OutLayout = ck::tensor_layout::convolution::GNHWK;
|
||||
|
||||
const auto in_g_n_c_wis_desc =
|
||||
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
|
||||
|
||||
const auto wei_g_k_c_xs_desc =
|
||||
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
|
||||
|
||||
// TODO - make_bias_host_tensor_descriptor_g_n_k_wos_packed()
|
||||
const auto bias_g_k_desc = HostTensorDescriptor({conv_param.G_,
|
||||
conv_param.N_,
|
||||
conv_param.K_,
|
||||
conv_param.output_spatial_lengths_[0],
|
||||
conv_param.output_spatial_lengths_[1]},
|
||||
{
|
||||
conv_param.K_, // g
|
||||
0, // n
|
||||
1, // k
|
||||
0, // ho
|
||||
0 // wo
|
||||
});
|
||||
|
||||
const auto requant_scale_g_k_desc = bias_g_k_desc;
|
||||
|
||||
const auto out_g_n_k_wos_desc =
|
||||
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
|
||||
|
||||
std::cout << out_g_n_k_wos_desc << std::endl;
|
||||
|
||||
using deviceOp = DeviceGroupedConvNDFwdInstance<ndim_spatial,
|
||||
InLayout,
|
||||
WeiLayout,
|
||||
BiasLayout,
|
||||
RequantScaleLayout,
|
||||
OutLayout>;
|
||||
|
||||
return run_grouped_conv_fwd<ndim_spatial,
|
||||
InDataType,
|
||||
WeiDataType,
|
||||
OutDataType,
|
||||
InElementOp,
|
||||
WeiElementOp,
|
||||
OutElementOp,
|
||||
deviceOp>(do_verification,
|
||||
time_kernel,
|
||||
conv_param,
|
||||
in_g_n_c_wis_desc,
|
||||
wei_g_k_c_xs_desc,
|
||||
bias_g_k_desc,
|
||||
requant_scale_g_k_desc,
|
||||
out_g_n_k_wos_desc,
|
||||
in_element_op,
|
||||
wei_element_op,
|
||||
out_element_op);
|
||||
}
|
||||
@@ -11,6 +11,7 @@
|
||||
#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/utility/convolution_parameter.hpp"
|
||||
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
|
||||
@@ -163,26 +164,25 @@ bool run_grouped_conv_fwd(bool do_verification,
|
||||
// do Conv
|
||||
auto conv = DeviceConvNDFwdInstance{};
|
||||
auto invoker = conv.MakeInvoker();
|
||||
auto argument = conv.MakeArgument(
|
||||
in_device_buf.GetDeviceBuffer(),
|
||||
wei_device_buf.GetDeviceBuffer(),
|
||||
std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
|
||||
out_device_buf.GetDeviceBuffer(),
|
||||
a_g_n_c_wis_lengths,
|
||||
a_g_n_c_wis_strides,
|
||||
b_g_k_c_xs_lengths,
|
||||
b_g_k_c_xs_strides,
|
||||
std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d0_g_n_k_wos_lengths}},
|
||||
std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d0_g_n_k_wos_strides}},
|
||||
e_g_n_k_wos_lengths,
|
||||
e_g_n_k_wos_strides,
|
||||
conv_filter_strides,
|
||||
conv_filter_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads,
|
||||
in_element_op,
|
||||
wei_element_op,
|
||||
out_element_op);
|
||||
auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
|
||||
wei_device_buf.GetDeviceBuffer(),
|
||||
{bias_device_buf.GetDeviceBuffer()},
|
||||
out_device_buf.GetDeviceBuffer(),
|
||||
a_g_n_c_wis_lengths,
|
||||
a_g_n_c_wis_strides,
|
||||
b_g_k_c_xs_lengths,
|
||||
b_g_k_c_xs_strides,
|
||||
{d0_g_n_k_wos_lengths},
|
||||
{d0_g_n_k_wos_strides},
|
||||
e_g_n_k_wos_lengths,
|
||||
e_g_n_k_wos_strides,
|
||||
conv_filter_strides,
|
||||
conv_filter_dilations,
|
||||
input_left_pads,
|
||||
input_right_pads,
|
||||
in_element_op,
|
||||
wei_element_op,
|
||||
out_element_op);
|
||||
|
||||
if(!conv.IsSupportedArgument(argument))
|
||||
{
|
||||
@@ -235,8 +235,8 @@ bool run_grouped_conv_fwd(bool do_verification,
|
||||
|
||||
out_device_buf.FromDevice(out_device.mData.data());
|
||||
|
||||
pass &= ck::utils::check_err(
|
||||
out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
|
||||
pass &=
|
||||
ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
|
||||
}
|
||||
|
||||
return (pass ? 0 : 1);
|
||||
@@ -11,6 +11,7 @@
|
||||
#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/utility/convolution_parameter.hpp"
|
||||
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
|
||||
@@ -150,14 +151,14 @@ bool run_grouped_conv_fwd(bool do_verification,
|
||||
auto invoker = conv.MakeInvoker();
|
||||
auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
|
||||
wei_device_buf.GetDeviceBuffer(),
|
||||
std::array<const void*, 0>{},
|
||||
{},
|
||||
out_device_buf.GetDeviceBuffer(),
|
||||
a_g_n_c_wis_lengths,
|
||||
a_g_n_c_wis_strides,
|
||||
b_g_k_c_xs_lengths,
|
||||
b_g_k_c_xs_strides,
|
||||
std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
|
||||
std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
|
||||
{},
|
||||
{},
|
||||
e_g_n_k_wos_lengths,
|
||||
e_g_n_k_wos_strides,
|
||||
conv_filter_strides,
|
||||
@@ -213,8 +214,8 @@ bool run_grouped_conv_fwd(bool do_verification,
|
||||
|
||||
out_device_buf.FromDevice(out_device.mData.data());
|
||||
|
||||
pass &= ck::utils::check_err(
|
||||
out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
|
||||
pass &=
|
||||
ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
|
||||
}
|
||||
|
||||
return (pass ? 0 : 1);
|
||||
Reference in New Issue
Block a user