Skip to content

GitLab

Commit 7a3b49e5 authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into contraction

parents e07b3d8e d3051d75
Hide whitespace changes
Inline Side-by-side
Showing with 1047 additions and 285 deletions
profiler/include/profile_gemm_add_add_fastgelu_impl.hpp 0 → 100644
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/host_tensor/host_conv.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using DeviceGemmAddAddFastGeluPtr = ck::tensor_operation::device::DeviceGemmMultipleDPtr<
2,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddAddFastGelu>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename D1DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename D1Layout,
typename ELayout>
int profile_gemm_add_add_fastgelu_impl(int do_verification,
int init_method,
bool /*do_log*/,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideD0,
int StrideD1,
int StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<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<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_host_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 << "d0_m_n: " << d0_m_n.mDesc << std::endl;
std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_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});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-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});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddAddFastGelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmAddAddFastGeluPtr>
device_op_ptrs;
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<EDataType, half_t>)
{
if constexpr(is_same_v<ALayout, tensor_layout::gemm::RowMajor> &&
is_same_v<BLayout, tensor_layout::gemm::RowMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
device_op_ptrs);
}
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::RowMajor> &&
is_same_v<BLayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
device_op_ptrs);
}
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<BLayout, tensor_layout::gemm::RowMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
device_op_ptrs);
}
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<BLayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
device_op_ptrs);
}
}
std::cout << "found " << device_op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n(HostTensorDescriptor(
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
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), d0_m_n(m, n), d1_m_n(m, n));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpace());
DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpace());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
std::string best_device_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
bool pass = true;
// profile device operation instances
for(auto& device_op_ptr : device_op_ptrs)
{
auto argument_ptr = device_op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
d1_m_n_device_buf.GetDeviceBuffer()},
static_cast<EDataType*>(e_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = device_op_ptr->MakeInvokerPointer();
std::string device_op_name = device_op_ptr->GetTypeString();
if(device_op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init E to zero before profiling a kernel
e_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), 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: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << device_op_name << std::endl;
if(tflops > best_tflops)
{
best_device_op_name = device_op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass &&
ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
}
}
else
{
std::cout << device_op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_device_op_name << std::endl;
return pass ? 0 : 1;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_gemm_bias_2d_impl.hpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm_bias.hpp"
#include "reference_gemm_bias_2d.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_bias.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp"
namespace ck {
namespace tensor_operation {
......
profiler/include/profile_gemm_bias_add_reduce_impl.hpp 0 → 100644
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmBiasAddReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmBiasAddReducePtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
DInElementOps,
DOutElementOps>;
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename C0DataType,
typename C1DataType,
typename DDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_bias_add_reduce_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int StrideC1)
{
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}));
};
auto f_host_tensor_descriptor2d =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<C0DataType> bias_n(f_host_tensor_descriptor1d(N, 1));
Tensor<C1DataType> c1_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_host_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
Tensor<DDataType> d0_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
Tensor<DDataType> d1_m_device_result(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
case 1:
std::srand(0);
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
bias_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
c1_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
break;
default:
std::srand(0);
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
bias_n.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5}, num_thread);
c1_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
using C1ElementOp = PassThrough;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto c1_element_op = C1ElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{N, N};
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
DDataType,
AElementOp,
BElementOp,
CElementOp>;
using ReduceAccDataType = DDataType;
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, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
for(int n = 0; n < N; ++n)
{
ReduceAccDataType acc = static_cast<ReduceAccDataType>(c_m_n_host_result(m, n)) +
static_cast<ReduceAccDataType>(bias_n(n));
ReduceAccDataType c1 = static_cast<ReduceAccDataType>(c1_m_n(m, n));
c_element_op(acc, acc);
c1_element_op(c1, c1);
acc += c1;
c_m_n_host_result(m, n) = static_cast<CDataType>(acc);
}
for(int m = 0; m < M; ++m)
{
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
ReduceAccDataType c_val =
ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
DeviceMem bias_device_buf(sizeof(C0DataType) * bias_n.mDesc.GetElementSpace());
DeviceMem c1_device_buf(sizeof(C1DataType) * c1_m_n.mDesc.GetElementSpace());
DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
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());
c1_device_buf.ToDevice(c1_m_n.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmBiasAddReduceNoOpPtr>
gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
{
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
gemm_ptrs);
}
}
if(gemm_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device GEMM instances
for(auto& gemm_ptr : gemm_ptrs)
{
auto argument_ptr = gemm_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<C0DataType*>(bias_device_buf.GetDeviceBuffer()),
static_cast<C1DataType*>(c1_device_buf.GetDeviceBuffer()),
&dxs_global,
M,
N,
K,
StrideA,
StrideB,
StrideC,
StrideC1,
a_element_op,
b_element_op,
c_element_op,
c1_element_op,
dxs_in_element_op,
dxs_out_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::string gemm_name = gemm_ptr->GetTypeString();
std::size_t flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
std::size_t num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N + sizeof(C0DataType) * M * N +
sizeof(C1DataType) * M * N + sizeof(DDataType) * M +
sizeof(DDataType) * M;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
d0_device_buf.FromDevice(d0_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_m_device_result.mData.data());
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
ck::utils::check_err(d0_m_device_result.mData, d0_m_host_result.mData);
ck::utils::check_err(d1_m_device_result.mData, d1_m_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host: ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_host: ", d0_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d0_device: ", d0_m_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_host: ", d1_m_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "d1_device: ", d1_m_device_result.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_gemm_bias_relu_add_impl.hpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm_bias_activation_add.hpp"
#include "reference_gemm_bias_activation_add.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_bias_activation_add.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation_add.hpp"
namespace ck {
namespace tensor_operation {
......
profiler/include/profile_gemm_bias_relu_impl.hpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm_bias_activation.hpp"
#include "reference_gemm_bias_activation.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_bias_activation.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm_bias_activation.hpp"
namespace ck {
namespace tensor_operation {
......
profiler/include/profile_gemm_impl.hpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include <iostream>
#include <typeinfo>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm.hpp"
#include "reference_gemm.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace tensor_operation {
......
profiler/include/profile_gemm_reduce_impl.hpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "reduction_operator.hpp"
#include "device_gemm_reduce.hpp"
#include "reference_gemm.hpp"
#include "ck/ck.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace tensor_operation {
......@@ -19,14 +24,13 @@ namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using Div = ck::tensor_operation::element_wise::UnaryDivide;
using Identity = ck::tensor_operation::element_wise::PassThrough;
using Square = ck::tensor_operation::element_wise::UnarySquare;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
DPtrsGlobal,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
......@@ -123,18 +127,16 @@ bool profile_gemm_reduce_impl(int do_verification,
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryIdentic<float, float, true>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add;
using D1ReduceOp = ck::reduce::Add;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryDivide;
using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
......@@ -143,7 +145,7 @@ bool profile_gemm_reduce_impl(int do_verification,
const auto d1_reduce_op = D1ReduceOp{};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{M, M};
auto dxs_out_element_op = DxsOutElementOps{N, N};
if(do_verification)
{
......@@ -155,6 +157,8 @@ bool profile_gemm_reduce_impl(int do_verification,
BElementOp,
CElementOp>;
using ReduceAccDataType = DDataType;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
......@@ -165,14 +169,15 @@ bool profile_gemm_reduce_impl(int do_verification,
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReductionZeroVal();
auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
for(int n = 0; n < N; ++n)
{
float c_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d0_val = 0;
float d1_val = 0;
ReduceAccDataType c_val =
ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
ReduceAccDataType d0_val;
ReduceAccDataType d1_val;
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
......@@ -257,7 +262,7 @@ bool profile_gemm_reduce_impl(int do_verification,
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
dxs_global,
&dxs_global,
M,
N,
K,
......@@ -309,13 +314,9 @@ bool profile_gemm_reduce_impl(int do_verification,
d0_device_buf.FromDevice(d0_m_device_result.mData.data());
d1_device_buf.FromDevice(d1_m_device_result.mData.data());
float c_error = check_error(c_m_n_host_result, c_m_n_device_result);
float d0_error = check_error(d0_m_host_result, d0_m_device_result);
float d1_error = check_error(d1_m_host_result, d1_m_device_result);
pass = pass && (c_error < 1E-6);
pass = pass && (d0_error < 1E-6);
pass = pass && (d1_error < 1E-6);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
ck::utils::check_err(d0_m_device_result.mData, d0_m_host_result.mData);
ck::utils::check_err(d1_m_device_result.mData, d1_m_host_result.mData);
if(do_log)
{
......
profiler/include/profile_grouped_gemm_impl.hpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "device_gemm.hpp"
#include "reference_gemm.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace tensor_operation {
......
profiler/include/profile_reduce_impl.hpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "check_err.hpp"
#include "device_reduce.hpp"
#include "device_reduce_instance.hpp"
#include "reduction_enums.hpp"
#include "host_reduction.hpp"
#include "host_common_util.hpp"
#include "host_tensor_generator.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_reduction.hpp"
#include "ck/library/host_tensor/host_common_util.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
namespace ck {
namespace tensor_operation {
......@@ -138,7 +143,6 @@ bool profile_reduce_impl_impl(bool do_verification,
{
using namespace ck::tensor_operation::device;
using namespace ck::tensor_operation::device::device_reduce_instance;
using namespace ck::host_reduce;
using ck::host_common::dumpBufferToFile;
constexpr bool op_support_indices =
......@@ -261,15 +265,22 @@ bool profile_reduce_impl_impl(bool do_verification,
float best_avg_time = 0;
float best_gb_per_sec = 0;
using InElementwiseOperation_0 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
InElementwiseOperation;
using AccElementwiseOperation_0 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
AccElementwiseOperation;
using InElementwiseOperation =
typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation =
typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
InElementwiseOperation in_elementwise_op;
AccElementwiseOperation acc_elementwise_op;
std::tie(in_elementwise_op, acc_elementwise_op) =
reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
static_cast<int32_t>(reduce_total_length));
using DeviceReduceInstPtr0 =
DeviceReducePtr<InElementwiseOperation_0, AccElementwiseOperation_0>;
DeviceReducePtr<InElementwiseOperation, AccElementwiseOperation>;
std::vector<DeviceReduceInstPtr0> reduce0_ptrs;
......@@ -313,15 +324,22 @@ bool profile_reduce_impl_impl(bool do_verification,
ReductionHost<InDataType,
AccDataType,
OutDataType,
ReduceOpId,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
Rank,
NumReduceDim,
PropagateNan,
OutputIndex>
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(
alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data());
hostReduce.Run(alpha,
in.mData.data(),
beta,
out_ref.mData.data(),
out_indices_ref.mData.data(),
in_elementwise_op,
acc_elementwise_op);
};
std::vector<ck::index_t> i_inLengths;
......@@ -336,11 +354,6 @@ bool profile_reduce_impl_impl(bool do_verification,
for(auto& reduce_ptr : reduce0_ptrs)
{
InElementwiseOperation_0 in_elementwise_op_0(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_0 acc_elementwise_op_0(
static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
......@@ -352,8 +365,8 @@ bool profile_reduce_impl_impl(bool do_verification,
nullptr,
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
in_elementwise_op_0,
acc_elementwise_op_0);
in_elementwise_op,
acc_elementwise_op);
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue;
......
profiler/src/profile_batched_gemm.cpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdint>
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_base.hpp"
#include "device_batched_gemm_xdl.hpp"
#include "profile_batched_gemm_impl.hpp"
#include "profiler/include/profile_batched_gemm_impl.hpp"
enum struct GemmMatrixLayout
{
......
profiler/src/profile_batched_gemm_reduce.cpp
View file @ 7a3b49e5
// 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 <stdlib.h>
#include <half.hpp>
#include "profile_batched_gemm_reduce_impl.hpp"
#include "profiler/include/profile_batched_gemm_reduce_impl.hpp"
int profile_batched_gemm_reduce(int argc, char* argv[])
{
......
profiler/src/profile_conv_bwd_weight.cpp
View file @ 7a3b49e5
// 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 <stdlib.h>
#include <half.hpp>
#include "profile_conv_bwd_weight_impl.hpp"
#include "profiler/include/profile_conv_bwd_weight_impl.hpp"
enum struct ConvDataType
{
......
profiler/src/profile_conv_fwd_bias_relu.cpp
View file @ 7a3b49e5
// 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 <stdlib.h>
#include <half.hpp>
#include "profile_conv_fwd_bias_relu_impl.hpp"
#include "profiler/include/profile_conv_fwd_bias_relu_impl.hpp"
enum struct ConvDataType
{
......
profiler/src/profile_conv_fwd_bias_relu_add.cpp
View file @ 7a3b49e5
// 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 <stdlib.h>
#include <half.hpp>
#include "profile_conv_fwd_bias_relu_add_impl.hpp"
#include "profiler/include/profile_conv_fwd_bias_relu_add_impl.hpp"
enum struct ConvDataType
{
......
profiler/src/profile_conv_fwd_bias_relu_atomic_add.cpp deleted 100644 → 0
View file @ e07b3d8e
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "profile_conv_fwd_bias_relu_atomic_add_impl.hpp"
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
enum struct ConvInputLayout
{
NCHW, // 0
NHWC, // 1
};
enum struct ConvWeightLayout
{
KCYX, // 0
KYXC, // 1
};
enum struct ConvOutputLayout
{
NKHW, // 0
NHWK, // 1
};
int profile_conv_fwd_bias_relu_atomic_add(int argc, char* argv[])
{
if(argc != 25)
{
printf("arg1: tensor operation (conv_fwd_bias_relu_atomic_add: "
"ForwardConvolution+Bias+ReLu+AtomicAdd)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
printf("arg6: verification (0: no; 1: yes)\n");
printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg9: time kernel (0=n0, 1=yes)\n");
printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
}
const auto data_type = static_cast<ConvDataType>(std::stoi(argv[2]));
const auto in_layout = static_cast<ConvInputLayout>(std::stoi(argv[3]));
const auto wei_layout = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
const auto out_layout = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const bool time_kernel = std::stoi(argv[9]);
const ck::index_t N = std::stoi(argv[10]);
const ck::index_t K = std::stoi(argv[11]);
const ck::index_t C = std::stoi(argv[12]);
const ck::index_t Y = std::stoi(argv[13]);
const ck::index_t X = std::stoi(argv[14]);
const ck::index_t Hi = std::stoi(argv[15]);
const ck::index_t Wi = std::stoi(argv[16]);
const ck::index_t conv_stride_h = std::stoi(argv[17]);
const ck::index_t conv_stride_w = std::stoi(argv[18]);
const ck::index_t conv_dilation_h = std::stoi(argv[19]);
const ck::index_t conv_dilation_w = std::stoi(argv[20]);
const ck::index_t in_left_pad_h = std::stoi(argv[21]);
const ck::index_t in_left_pad_w = std::stoi(argv[22]);
const ck::index_t in_right_pad_h = std::stoi(argv[23]);
const ck::index_t in_right_pad_w = std::stoi(argv[24]);
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_fwd_bias_relu_atomic_add_impl<
2,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
time_kernel,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else
{
throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
}
return 0;
}
profiler/src/profile_convnd_bwd_data.cpp
View file @ 7a3b49e5
// 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 <stdlib.h>
#include <half.hpp>
#include "profile_convnd_bwd_data_impl.hpp"
#include "profiler/include/profile_convnd_bwd_data_impl.hpp"
namespace {
......
profiler/src/profile_convnd_fwd.cpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <functional>
#include <iostream>
#include <memory>
#include <string>
#include <vector>
#include <half.hpp>
#include "conv_util.hpp"
#include "element_wise_operation.hpp"
#include "fill.hpp"
#include "profile_convnd_fwd.hpp"
#include "tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/utility/fill.hpp"
#include "profiler/include/profile_convnd_fwd.hpp"
namespace {
......@@ -150,9 +154,12 @@ void profile_convnd_instances_impl(const ck::utils::conv::ConvParams& params,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::utils::FillUniform<int>,
ck::utils::FillUniform<int>>>(
params, true, ck::utils::FillUniform<int>{}, ck::utils::FillUniform<int>{});
ck::utils::FillUniformDistributionIntegerValue<int>,
ck::utils::FillUniformDistributionIntegerValue<int>>>(
params,
true,
ck::utils::FillUniformDistributionIntegerValue<int>{},
ck::utils::FillUniformDistributionIntegerValue<int>{});
break;
case 2:
conv_instance = std::make_unique<
......@@ -165,12 +172,12 @@ void profile_convnd_instances_impl(const ck::utils::conv::ConvParams& params,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::utils::FillUniform<InDataType>,
ck::utils::FillUniform<WeiDataType>>>(
ck::utils::FillUniformDistribution<InDataType>,
ck::utils::FillUniformDistribution<WeiDataType>>>(
params,
true,
ck::utils::FillUniform<InDataType>{},
ck::utils::FillUniform<WeiDataType>{});
ck::utils::FillUniformDistribution<InDataType>{},
ck::utils::FillUniformDistribution<WeiDataType>{});
break;
default: throw std::runtime_error("Unsupported init method!");
}
......@@ -181,8 +188,10 @@ void profile_convnd_instances_impl(const ck::utils::conv::ConvParams& params,
_1,
_2,
_3);
OpInstanceRunEngine<InDataType, WeiDataType, OutDataType> run_engine(*conv_instance,
reference_conv_fwd_fun);
OpInstanceRunEngine<InDataType, WeiDataType, OutDataType> run_engine(
*conv_instance, reference_conv_fwd_fun, do_verification);
auto best_conf = run_engine.Profile(
conv::ConvolutionFwdInstances<InDataType, WeiDataType, OutDataType>::template Get<NDim>(),
time_kernel,
......
profiler/src/profile_gemm.cpp
View file @ 7a3b49e5
// 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 <stdlib.h>
#include <half.hpp>
#include "profile_gemm_impl.hpp"
#include "profiler/include/profile_gemm_impl.hpp"
enum struct GemmMatrixLayout
{
......
profiler/src/profile_gemm_add_add_fastgelu.cpp 0 → 100644
View file @ 7a3b49e5
// 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 "profiler/include/profile_gemm_add_add_fastgelu_impl.hpp"
int profile_gemm_add_add_fastgelu(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN_MN, // 0
MK_NK_MN_MN_MN, // 1
KM_KN_MN_MN_MN, // 2
KM_NK_MN_MN_MN, // 3
MK_KN_NM_MN_MN, // 4
MK_NK_NM_MN_MN, // 5
KM_KN_NM_MN_MN, // 6
KM_NK_NM_MN_MN, // 7
};
enum struct MatrixDataType
{
F32_F32_F32_F32_F32, // 0
F16_F16_F16_F16_F16, // 1
BF16_BF16_BF16_BF16_BF16, // 2
INT8_INT8_INT8_INT8_INT8, // 3
};
if(argc != 16)
{
// clang-format off
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+GeLU)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
printf(" 1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
printf(" 2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n] + D1[m, n]);\n");
printf(" 3: E[m, n] = FastGeLU(A[k, m] * B[n, k] + D0[m, n] + D1[m, n]))\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
// clang-format on
exit(1);
}
const auto data_type = static_cast<MatrixDataType>(std::stoi(argv[2]));
const auto layout = static_cast<MatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideD0 = std::stoi(argv[13]);
const int StrideD1 = std::stoi(argv[14]);
const int StrideE = std::stoi(argv[15]);
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto d0_type,
auto d1_type,
auto e_type,
auto a_layout,
auto b_layout,
auto d0_layout,
auto d1_layout,
auto e_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using D0DataType = decltype(d0_type);
using D1DataType = decltype(d1_type);
using EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_layout);
using D1Layout = decltype(d1_layout);
using ELayout = decltype(e_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
return ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
D1DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
D1Layout,
ELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
(StrideE < 0) ? DefaultStrideE : StrideE);
};
if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::MK_NK_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::KM_KN_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::KM_NK_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 0;
}
}
profiler/src/profile_gemm_bias_2d.cpp
View file @ 7a3b49e5
// 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 <stdlib.h>
#include <half.hpp>
#include "profile_gemm_bias_2d_impl.hpp"
#include "profiler/include/profile_gemm_bias_2d_impl.hpp"
enum struct GemmMatrixLayout
{
......
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