Skip to content

GitLab

Commit b30d416c authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'develop' into amd-develop

parents 2fd6c6d4 94fbaac0
Hide whitespace changes
Inline Side-by-side
Showing with 1406 additions and 298 deletions
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
View file @ b30d416c
......@@ -2,7 +2,7 @@
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_i8_instance.hpp"
namespace ck {
namespace tensor_operation {
......
profiler/include/profiler/profile_gemm_add_impl.hpp 0 → 100644
View file @ b30d416c
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, 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/tensor_operation_instance/gpu/gemm_add.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/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout>
bool profile_gemm_add_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 StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), 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<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
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 << "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});
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});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Add = ck::tensor_operation::element_wise::Add;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Add>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, 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));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_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());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
std::string best_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& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(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, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = 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, e_m_n_host_result);
}
}
else
{
std::cout << 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_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_gemm_add_relu_impl.hpp 0 → 100644
View file @ b30d416c
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, 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/tensor_operation_instance/gpu/gemm_add_relu.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/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout>
bool profile_gemm_add_relu_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 StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), 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<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
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 << "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});
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});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddRelu = ck::tensor_operation::element_wise::AddRelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddRelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddRelu>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, 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));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_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());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
std::string best_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& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(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, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = 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, e_m_n_host_result);
}
}
else
{
std::cout << 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_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_gemm_add_silu_impl.hpp 0 → 100644
View file @ b30d416c
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, 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/tensor_operation_instance/gpu/gemm_add_silu.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/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout>
bool profile_gemm_add_silu_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 StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), 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<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
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 << "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});
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});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddRelu = ck::tensor_operation::element_wise::AddSilu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddRelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddSilu>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, 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));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_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());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
std::string best_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& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(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, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = 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, e_m_n_host_result);
}
}
else
{
std::cout << 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_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/src/CMakeLists.txt
View file @ b30d416c
......@@ -43,7 +43,10 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
......@@ -109,7 +112,10 @@ if(DL_KERNELS)
endif()
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
......
profiler/src/profile_gemm_add.cpp 0 → 100644
View file @ b30d416c
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_add_impl.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "gemm_add"
#define OP_DESC "GEMM+Add"
using INT8 = int8_t;
using BF16 = ck::bhalf_t;
int profile_gemm_add(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN, // 0
MK_NK_MN_MN, // 1
KM_KN_MN_MN, // 2
KM_NK_MN_MN, // 3
};
enum struct MatrixDataType
{
F16_INT8_F16_F16, // 0
BF16_INT8_BF16_BF16, // 1
};
if(argc != 15)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: f16&i8 1: bf16&i8)\n");
printf("arg3: matrix layout (0: E[m, n] = ReLU(A[m, k] * B[k, n] + D0[m, n]);\n");
printf(" 1: E[m, n] = ReLU(A[m, k] * B[n, k] + D0[m, n]);\n");
printf(" 2: E[m, n] = ReLU(A[k, m] * B[k, n] + D0[m, n]);\n");
printf(" 3: E[m, n] = ReLU(A[k, m] * B[n, k] + D0[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 14: M, N, K, StrideA, StrideB, StrideD0, 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 StrideE = std::stoi(argv[14]);
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 e_type,
auto a_layout,
auto b_layout,
auto d0_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 EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_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 DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_add_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideE < 0) ? DefaultStrideE : StrideE);
return pass ? 0 : 1;
};
if(data_type == MatrixDataType::F16_INT8_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(F16{}, INT8{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::BF16_INT8_BF16_BF16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(BF16{}, INT8{}, F32{}, BF16{}, BF16{}, Row{}, Row{}, Row{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add);
profiler/src/profile_gemm_add_fastgelu.cpp
View file @ b30d416c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
......@@ -12,6 +12,9 @@
#define OP_NAME "gemm_add_fastgelu"
#define OP_DESC "GEMM+Add+FastGeLU"
using INT8 = int8_t;
using BF16 = ck::bhalf_t;
int profile_gemm_add_fastgelu(int argc, char* argv[])
{
enum struct MatrixLayout
......@@ -28,13 +31,15 @@ int profile_gemm_add_fastgelu(int argc, char* argv[])
F16_F16_F16_F16, // 1
BF16_BF16_BF16_BF16, // 2
INT8_INT8_INT8_INT8, // 3
F16_INT8_F16_F16, // 4
BF16_INT8_BF16_BF16, // 5
};
if(argc != 15)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f16&i8 5: bf16&i8)\n");
printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n]);\n");
printf(" 1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n]);\n");
printf(" 2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n]);\n");
......@@ -135,6 +140,14 @@ int profile_gemm_add_fastgelu(int argc, char* argv[])
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_INT8_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(F16{}, INT8{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::BF16_INT8_BF16_BF16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(BF16{}, INT8{}, F32{}, BF16{}, BF16{}, Row{}, Row{}, Row{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
......
profiler/src/profile_gemm_add_relu.cpp 0 → 100644
View file @ b30d416c
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_add_relu_impl.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "gemm_add_relu"
#define OP_DESC "GEMM+Add+ReLU"
using INT8 = int8_t;
using BF16 = ck::bhalf_t;
int profile_gemm_add_relu(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN, // 0
MK_NK_MN_MN, // 1
KM_KN_MN_MN, // 2
KM_NK_MN_MN, // 3
};
enum struct MatrixDataType
{
F16_INT8_F16_F16, // 0
BF16_INT8_BF16_BF16, // 1
};
if(argc != 15)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: f16&i8 1: bf16&i8)\n");
printf("arg3: matrix layout (0: E[m, n] = ReLU(A[m, k] * B[k, n] + D0[m, n]);\n");
printf(" 1: E[m, n] = ReLU(A[m, k] * B[n, k] + D0[m, n]);\n");
printf(" 2: E[m, n] = ReLU(A[k, m] * B[k, n] + D0[m, n]);\n");
printf(" 3: E[m, n] = ReLU(A[k, m] * B[n, k] + D0[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 14: M, N, K, StrideA, StrideB, StrideD0, 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 StrideE = std::stoi(argv[14]);
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 e_type,
auto a_layout,
auto b_layout,
auto d0_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 EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_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 DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_add_relu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideE < 0) ? DefaultStrideE : StrideE);
return pass ? 0 : 1;
};
if(data_type == MatrixDataType::F16_INT8_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(F16{}, INT8{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::BF16_INT8_BF16_BF16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(BF16{}, INT8{}, F32{}, BF16{}, BF16{}, Row{}, Row{}, Row{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_relu);
profiler/src/profile_gemm_add_silu.cpp 0 → 100644
View file @ b30d416c
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/profile_gemm_add_silu_impl.hpp"
#include "profiler_operation_registry.hpp"
#define OP_NAME "gemm_add_silu"
#define OP_DESC "GEMM+Add+SiLU"
using INT8 = int8_t;
using BF16 = ck::bhalf_t;
int profile_gemm_add_silu(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN, // 0
MK_NK_MN_MN, // 1
KM_KN_MN_MN, // 2
KM_NK_MN_MN, // 3
};
enum struct MatrixDataType
{
F16_INT8_F16_F16, // 0
BF16_INT8_BF16_BF16, // 1
};
if(argc != 15)
{
// clang-format off
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: f16&i8 1: bf16&i8)\n");
printf("arg3: matrix layout (0: E[m, n] = ReLU(A[m, k] * B[k, n] + D0[m, n]);\n");
printf(" 1: E[m, n] = ReLU(A[m, k] * B[n, k] + D0[m, n]);\n");
printf(" 2: E[m, n] = ReLU(A[k, m] * B[k, n] + D0[m, n]);\n");
printf(" 3: E[m, n] = ReLU(A[k, m] * B[n, k] + D0[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 14: M, N, K, StrideA, StrideB, StrideD0, 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 StrideE = std::stoi(argv[14]);
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 e_type,
auto a_layout,
auto b_layout,
auto d0_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 EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_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 DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_add_silu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideE < 0) ? DefaultStrideE : StrideE);
return pass ? 0 : 1;
};
if(data_type == MatrixDataType::F16_INT8_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(F16{}, INT8{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::BF16_INT8_BF16_BF16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(BF16{}, INT8{}, F32{}, BF16{}, BF16{}, Row{}, Row{}, Row{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_silu);
script/parse_perf_data.py deleted 100644 → 0
View file @ 2fd6c6d4
#!/usr/bin/env python3
import os, io, argparse, datetime, re
import numpy as np
import sqlalchemy
from sqlalchemy.types import NVARCHAR, Float, Integer
import pymysql
import pandas as pd
from sshtunnel import SSHTunnelForwarder
def print_to_string(*args, **kwargs):
output = io.StringIO()
print(*args, file=output, **kwargs)
contents = output.getvalue()
output.close()
return contents
def parse_args():
parser = argparse.ArgumentParser(description='Parse results from tf benchmark runs')
parser.add_argument('filename', type=str, help='Log file to prase or directory containing log files')
args = parser.parse_args()
files = []
if os.path.isdir(args.filename):
all_files = os.listdir(args.filename)
for name in all_files:
if not 'log' in name:
continue
files.append(os.path.join(args.filename, name))
else:
files = [args.filename]
args.files = files
return args
def main():
args = parse_args()
tests = []
kernels=[]
tflops=[]
dtype=[]
alayout=[]
blayout=[]
M=[]
N=[]
K=[]
StrideA=[]
StrideB=[]
StrideC=[]
#parse results, get the Tflops value for "Best Perf" kernels
glue=""
for filename in args.files:
for line in open(filename):
if 'Branch name' in line:
lst=line.split()
branch_name=lst[2]
if 'On branch' in line:
lst=line.split()
branch_name=lst[2]
if 'Node name' in line:
lst=line.split()
node_id=lst[2]
if 'GPU_arch' in line:
lst=line.split()
gpu_arch=lst[2]
if 'HIP version' in line:
lst=line.split()
hip_vers=lst[2]
if 'Compute Unit' in line:
lst=line.split()
compute_units=lst[2]
if 'InstalledDir' in line:
lst=line.split()
rocm_vers=lst[1][lst[1].find('/opt/rocm-')+len('/opt/rocm-'):lst[1].rfind('/llvm/bin')]
print("Branch name:",branch_name)
print("Node name:",node_id)
print("GPU_arch:",gpu_arch)
print("Compute units:",compute_units)
print("ROCM_version:",rocm_vers)
print("HIP_version:",hip_vers)
#parse gemm performance tests:
if 'gemm' in filename:
for filename in args.files:
for line in open(filename):
if 'Best Perf' in line:
lst=line.split()
if len(lst)>=37: #the line is complete
tests.append(glue.join(lst[5:30]))
kernels.append(glue.join(lst[37:]))
tflops.append(lst[33])
dtype.append(lst[5])
alayout.append(lst[8])
blayout.append(lst[11])
M.append(lst[14])
N.append(lst[17])
K.append(lst[20])
StrideA.append(lst[23])
StrideB.append(lst[26])
StrideC.append(lst[29])
elif len(lst)<37 and len(lst)>=33: #the tflops are available
tests.append(glue.join(lst[5:30]))
kernels.append("N/A")
tflops.append(lst[33])
dtype.append(lst[5])
alayout.append(lst[8])
blayout.append(lst[11])
M.append(lst[14])
N.append(lst[17])
K.append(lst[20])
StrideA.append(lst[23])
StrideB.append(lst[26])
StrideC.append(lst[29])
print("warning: incomplete line:",lst)
elif len(lst)<33: #even the tflops are not available
print("Error in ckProfiler output!")
print("warning: incomplete line=",lst)
#sort results
#sorted_tests = sorted(tests)
#print("sorted tests:",sorted_tests)
sorted_tflops = [x for _,x in sorted(zip(tests,tflops))]
#sorted_kernels = [x for _,x in sorted(zip(tests,kernels))]
test_list=list(range(1,len(tests)+1))
#parse resnet50 performance tests:
if 'resnet50' in filename:
for filename in args.files:
for line in open(filename):
if 'Best Perf' in line:
lst=line.split()
tflops.append(lst[4])
print("Number of tests:",len(tflops))
sql_hostname = '127.0.0.1'
sql_username = os.environ["dbuser"]
sql_password = os.environ["dbpassword"]
sql_main_database = 'miopen_perf'
sql_port = 3306
ssh_host = os.environ["dbsship"]
ssh_user = os.environ["dbsshuser"]
ssh_port = int(os.environ["dbsshport"])
ssh_pass = os.environ["dbsshpassword"]
with SSHTunnelForwarder(
(ssh_host, ssh_port),
ssh_username=ssh_user,
ssh_password=ssh_pass,
remote_bind_address=(sql_hostname, sql_port)) as tunnel:
sqlEngine = sqlalchemy.create_engine('mysql+pymysql://{0}:{1}@{2}:{3}/{4}'.
format(sql_username, sql_password, sql_hostname, tunnel.local_bind_port, sql_main_database))
conn = sqlEngine.connect()
#save gemm performance tests:
if 'gemm' in filename:
#write the ck_gemm_test_params table
#only needed once the test set changes
'''
sorted_dtypes = [x for _,x in sorted(zip(tests,dtype))]
sorted_alayout = [x for _,x in sorted(zip(tests,alayout))]
sorted_blayout = [x for _,x in sorted(zip(tests,blayout))]
sorted_M = [x for _,x in sorted(zip(tests,M))]
sorted_N = [x for _,x in sorted(zip(tests,N))]
sorted_K = [x for _,x in sorted(zip(tests,K))]
sorted_StrideA = [x for _,x in sorted(zip(tests,StrideA))]
sorted_StrideB = [x for _,x in sorted(zip(tests,StrideB))]
sorted_StrideC = [x for _,x in sorted(zip(tests,StrideC))]
ck_gemm_params=[test_list,sorted_dtypes,sorted_alayout,sorted_blayout,
sorted_M,sorted_N,sorted_K,sorted_StrideA,sorted_StrideB,
sorted_StrideC]
df=pd.DataFrame(np.transpose(ck_gemm_params),columns=['Test_number','Data_type',
'Alayout','BLayout','M','N','K', 'StrideA','StrideB','StrideC'])
print(df)
dtypes = {
'Test_number': Integer(),
'Data_type': NVARCHAR(length=5),
'Alayout': NVARCHAR(length=12),
'Blayout': NVARCHAR(length=12),
'M': Integer(),
'N': Integer(),
'K': Integer(),
'StrideA': Integer(),
'StrideB': Integer(),
'StrideC': Integer()
}
df.to_sql("ck_gemm_test_params",conn,if_exists='replace',index=False, dtype=dtypes)
'''
#read baseline results for the latest develop branch
query = '''SELECT * from ck_gemm_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_gemm_tflops where Branch_ID='develop' );'''
tflops_base = pd.read_sql_query(query, conn)
#write new results to the db
testlist=[]
for i in range(1,len(tests)+1):
testlist.append("Test%i"%i)
ck_gemm_tflops=[str(branch_name),str(node_id),str(gpu_arch),compute_units,str(rocm_vers),str(hip_vers),str(datetime.datetime.now())]
flops=pd.DataFrame(data=[ck_gemm_tflops],columns=['Branch_ID','Node_ID','GPU_arch','Compute Units','ROCM_version','HIP_version','Datetime'])
df_add=pd.DataFrame(data=[sorted_tflops],columns=testlist)
flops=pd.concat([flops,df_add],axis=1)
print("new tflops for gemm tests:",flops)
flops.to_sql("ck_gemm_tflops",conn,if_exists='append',index=False)
#save resnet50 performance tests:
if 'resnet50' in filename:
#read baseline results for the latest develop branch
query = '''SELECT * from ck_resnet50_N256_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_resnet50_N256_tflops where Branch_ID='develop' );'''
tflops_base_N256 = pd.read_sql_query(query, conn)
query = '''SELECT * from ck_resnet50_N4_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_resnet50_N4_tflops where Branch_ID='develop' );'''
tflops_base_N4 = pd.read_sql_query(query, conn)
#write new results to the db
testlist=[]
for i in range(1,50):
testlist.append("Layer%i"%i)
ck_resnet_tflops=[str(branch_name),str(node_id),str(gpu_arch),compute_units,str(rocm_vers),str(hip_vers),str(datetime.datetime.now())]
flops0=pd.DataFrame(data=[ck_resnet_tflops],columns=['Branch_ID','Node_ID','GPU_arch','Compute Units','ROCM_version','HIP_version','Datetime'])
df_add=pd.DataFrame(data=[tflops[0:49]],columns=testlist)
flops=pd.concat([flops0,df_add],axis=1)
print("new tflops for N=256 resnet50 test:",flops)
flops.to_sql("ck_resnet50_N256_tflops",conn,if_exists='append',index=False)
df_add=pd.DataFrame(data=[tflops[49:98]],columns=testlist)
flops=pd.concat([flops0,df_add],axis=1)
print("new tflops for N=4 resnet50 test:",flops)
flops.to_sql("ck_resnet50_N4_tflops",conn,if_exists='append',index=False)
conn.close()
#compare the results to the baseline if baseline exists
regression=0
if 'gemm' in filename:
if not tflops_base.empty:
base=tflops_base[testlist].to_numpy(dtype='float')
base_list=base[0]
ave_perf=0
for i in range(len(base_list)):
# success criterion:
if base_list[i]>1.01*float(sorted_tflops[i]):
print("test # ",i,"shows regression by {:.3f}%".format(
(float(sorted_tflops[i])-base_list[i])/base_list[i]*100))
regression=1
ave_perf=ave_perf+float(sorted_tflops[i])/base_list[i]
if regression==0:
print("no regressions found")
ave_perf=ave_perf/len(base_list)
print("average performance relative to baseline:",ave_perf)
else:
print("could not find a baseline")
if 'resnet50' in filename:
if not tflops_base_N256.empty:
base=tflops_base_N256[testlist].to_numpy(dtype='float')
base_list=base[0]
ave_perf=0
for i in range(len(base_list)):
# success criterion:
if base_list[i]>1.01*float(tflops[i]):
print("layer # ",i,"shows regression by {:.3f}%".format(
(float(tflops[i])-base_list[i])/base_list[i]*100))
regression=1
ave_perf=ave_perf+float(tflops[i])/base_list[i]
if regression==0:
print("no regressions found")
ave_perf=ave_perf/len(base_list)
print("average performance relative to baseline:",ave_perf)
else:
print("could not find a baseline for N=256")
if not tflops_base_N4.empty:
base=tflops_base_N4[testlist].to_numpy(dtype='float')
base_list=base[0]
ave_perf=0
for i in range(len(base_list)):
# success criterion:
if base_list[i]>1.01*float(tflops[i+49]):
print("layer # ",i,"shows regression by {:.3f}%".format(
(float(tflops[i+49])-base_list[i])/base_list[i]*100))
regression=1
ave_perf=ave_perf+float(tflops[i+49])/base_list[i]
if regression==0:
print("no regressions found")
ave_perf=ave_perf/len(base_list)
print("average performance relative to baseline:",ave_perf)
else:
print("could not find a baseline for N=4")
#return 0 if performance criteria met, otherwise return 1
return regression
if __name__ == '__main__':
main()
\ No newline at end of file
script/process_perf_data.py
View file @ b30d416c
......@@ -133,7 +133,7 @@ def parse_logfile(logfile):
if 'Best Perf' in line:
lst=line.split()
res.append(lst[4])
elif 'onnx_gemm' in logfile or 'splitK_gemm' in logfile:
elif 'onnx_gemm' in logfile or 'splitK_gemm' in logfile or 'mixed_gemm' in logfile:
for line in open(logfile):
if 'Best Perf' in line:
lst=line.split()
......@@ -295,6 +295,10 @@ def main():
for i in range(1,len(results)+1):
testlist.append("Test%i"%i)
table_name="ck_splitK_gemm_tflops"
if 'mixed_gemm' in filename:
for i in range(1,len(results)+1):
testlist.append("Test%i"%i)
table_name="ck_mixed_gemm_tflops"
tflops_base = get_baseline(table_name,conn)
store_new_test_result(table_name, results, testlist, branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment, conn)
......
script/profile_mixed_gemm.sh 0 → 100755
View file @ b30d416c
#!/bin/bash
## GPU visibility
export HIP_VISIBLE_DEVICES=0
DRIVER="../build/bin/ckProfiler"
echo $DRIVER
OP=$1
DATATYPE=$2
LAYOUT=$3
VERIFY=$4
INIT=$5
LOG=$6
TIME=$7
KBatch=$8
######## op datatype layout verify init log time M___ N___ K___ StrideA StrideB StrideC KBatch_
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 16 16 1024 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 16 16 8192 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 16 16 65536 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 16 2048 1024 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 16 2048 8192 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 16 2048 65536 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 16 8192 1024 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 16 8192 8192 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 16 8192 65536 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048 16 1024 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048 16 8192 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048 16 65536 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048 2048 1024 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048 2048 8192 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048 2048 65536 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048 8192 1024 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048 8192 8192 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048 8192 65536 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192 16 1024 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192 16 8192 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192 16 65536 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192 2048 1024 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192 2048 8192 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192 2048 65536 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192 8192 1024 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192 8192 8192 -1 -1 -1 $KBatch
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192 8192 65536 -1 -1 -1 $KBatch
\ No newline at end of file
script/run_full_performance_tests.sh
View file @ b30d416c
......@@ -147,3 +147,9 @@ export onnx_log="perf_onnx_gemm.log"
print_log_header $onnx_log $env_type $branch $host_name
./profile_onnx_gemm.sh gemm 0 0 $verify 1 0 1 2>&1 | tee -a $onnx_log
./profile_onnx_gemm.sh gemm 1 0 $verify 1 0 1 2>&1 | tee -a $onnx_log
#run mixed fp16/fp8 and fp8/fp16 gemm tests
export mixed_gemm_log="perf_mixed_gemm.log"
print_log_header $mixed_gemm_log $env_type $branch $host_name
./profile_mixed_gemm.sh gemm_splitk 4 0 $verify 2 0 1 16 2>&1 | tee -a $mixed_gemm_log
./profile_mixed_gemm.sh gemm_splitk 5 0 $verify 2 0 1 16 2>&1 | tee -a $mixed_gemm_log
\ No newline at end of file
test/CMakeLists.txt
View file @ b30d416c
......@@ -122,6 +122,7 @@ add_subdirectory(space_filling_curve)
add_subdirectory(conv_util)
add_subdirectory(reference_conv_fwd)
add_subdirectory(gemm)
add_subdirectory(gemm_add)
add_subdirectory(gemm_layernorm)
add_subdirectory(gemm_split_k)
add_subdirectory(gemm_reduce)
......
test/gemm_add/CMakeLists.txt 0 → 100644
View file @ b30d416c
add_gtest_executable(test_gemm_add test_gemm_add.hpp)
target_link_libraries(test_gemm_add PRIVATE utility device_gemm_add_instance)
add_gtest_executable(test_gemm_add_relu test_gemm_add_relu.cpp)
target_link_libraries(test_gemm_add_relu PRIVATE utility device_gemm_add_instance device_gemm_add_relu_instance)
add_gtest_executable(test_gemm_add_silu test_gemm_add_silu.cpp)
target_link_libraries(test_gemm_add_silu PRIVATE utility device_gemm_add_instance device_gemm_add_silu_instance)
add_gtest_executable(test_gemm_add_fastgelu test_gemm_add_fastgelu.cpp)
target_link_libraries(test_gemm_add_fastgelu PRIVATE utility device_gemm_add_instance device_gemm_add_fastgelu_instance)
test/gemm_add/test_gemm_add.hpp 0 → 100644
View file @ b30d416c
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/ck.hpp"
#include "profiler/profile_gemm_add_impl.hpp"
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using I8 = int8_t;
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
template <typename Tuple>
class TestGemmAdd : public ::testing::Test
{
protected:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
constexpr static auto ProfileGemmAddImpl = ck::profiler::profile_gemm_add_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
virtual decltype(ProfileGemmAddImpl) GetImpl() { return ProfileGemmAddImpl; }
void Run()
{
std::vector<std::vector<ck::index_t>> lengths = {
{16, 32, 64}, {2048, 4096, 8192}, {2048, 1024, 16}};
bool all_success = true;
for(auto length : lengths)
{
int M = length[0];
int N = length[1];
int K = length[2];
int StrideA = ck::is_same_v<ALayout, Row> ? K : M;
int StrideB = ck::is_same_v<BLayout, Row> ? N : K;
int StrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
int StrideE = ck::is_same_v<ELayout, Row> ? N : M;
all_success =
all_success &
GetImpl()(true, 1, false, false, M, N, K, StrideA, StrideB, StrideD0, StrideE);
}
EXPECT_TRUE(all_success);
}
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,
std::tuple<BF16, I8, F32, BF16, BF16, Row, Row, Row, Row>>;
TYPED_TEST_SUITE(TestGemmAdd, KernelTypes);
TYPED_TEST(TestGemmAdd, Test_BF16FP16_INT8) { this->Run(); }
test/gemm_add/test_gemm_add_fastgelu.cpp 0 → 100644
View file @ b30d416c
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/ck.hpp"
#include "profiler/profile_gemm_add_fastgelu_impl.hpp"
#include "test_gemm_add.hpp"
template <typename Tuple>
class TestGemmAddFastgelu : public TestGemmAdd<Tuple>
{
private:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
constexpr static auto ProfileGemmAddFastgeluImpl =
ck::profiler::profile_gemm_add_fastgelu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
decltype(ProfileGemmAddFastgeluImpl) GetImpl() override { return ProfileGemmAddFastgeluImpl; }
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,
std::tuple<BF16, I8, F32, BF16, BF16, Row, Row, Row, Row>>;
TYPED_TEST_SUITE(TestGemmAddFastgelu, KernelTypes);
TYPED_TEST(TestGemmAddFastgelu, Test_BF16FP16) { this->Run(); }
test/gemm_add/test_gemm_add_relu.cpp 0 → 100644
View file @ b30d416c
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/ck.hpp"
#include "profiler/profile_gemm_add_relu_impl.hpp"
#include "test_gemm_add.hpp"
template <typename Tuple>
class TestGemmAddRelu : public TestGemmAdd<Tuple>
{
private:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
constexpr static auto ProfileGemmAddReluImpl =
ck::profiler::profile_gemm_add_relu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
decltype(ProfileGemmAddReluImpl) GetImpl() override { return ProfileGemmAddReluImpl; }
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,
std::tuple<BF16, I8, F32, BF16, BF16, Row, Row, Row, Row>>;
TYPED_TEST_SUITE(TestGemmAddRelu, KernelTypes);
TYPED_TEST(TestGemmAddRelu, Test_BF16FP16_INT8) { this->Run(); }
test/gemm_add/test_gemm_add_silu.cpp 0 → 100644
View file @ b30d416c
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "ck/ck.hpp"
#include "profiler/profile_gemm_add_silu_impl.hpp"
#include "test_gemm_add.hpp"
template <typename Tuple>
class TestGemmAddSilu : public TestGemmAdd<Tuple>
{
private:
using ADataType = std::tuple_element_t<0, Tuple>;
using BDataType = std::tuple_element_t<1, Tuple>;
using AccDataType = std::tuple_element_t<2, Tuple>;
using D0DataType = std::tuple_element_t<3, Tuple>;
using EDataType = std::tuple_element_t<4, Tuple>;
using ALayout = std::tuple_element_t<5, Tuple>;
using BLayout = std::tuple_element_t<6, Tuple>;
using D0Layout = std::tuple_element_t<7, Tuple>;
using ELayout = std::tuple_element_t<8, Tuple>;
constexpr static auto ProfileGemmAddSiluImpl =
ck::profiler::profile_gemm_add_silu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
ELayout>;
decltype(ProfileGemmAddSiluImpl) GetImpl() override { return ProfileGemmAddSiluImpl; }
};
using KernelTypes = ::testing::Types<std::tuple<F16, I8, F32, F16, F16, Row, Row, Row, Row>,
std::tuple<BF16, I8, F32, BF16, BF16, Row, Row, Row, Row>>;
TYPED_TEST_SUITE(TestGemmAddSilu, KernelTypes);
TYPED_TEST(TestGemmAddSilu, Test_BF16FP16_INT8) { this->Run(); }
test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
View file @ b30d416c
......@@ -55,10 +55,7 @@ class TestGroupedConvndBwdWeight : public ::testing::Test
}
}
const bool is_navi3x = ck::get_device_name() == "gfx1100" ||
ck::get_device_name() == "gfx1101" ||
ck::get_device_name() == "gfx1102";
if(is_navi3x)
if(ck::is_navi3_supported())
{
// on navi3x only support for 3d is implemented
if constexpr(NDimSpatial{} != 3)
......
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