mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-26 08:00:13 +00:00
GEMM with Multiple Source, GEMM+Bias+Add+FastGeLU example and ckProfiler (#241)
* ad gelu and fast_gelu
* added GeLU and fast GeLU
* clean up
* add gemm+fastgelu example
* add gemm+gelu instances
* update profiler
* clean up
* clean up
* adding gemm+bias+activation
* clean
* adding bias
* clean
* adding gemm multiple d
* debugging
* add gemm bias add fastgelu
* rename, clean
* refactoring; add readme
* refactor
* refactor
* refactor
* refactor
* refactor
* refactor
* fix
* fix
* update example
* update example
* rename
* update example
* add ckProfiler
* clean
* clean
* clean
* clean
* add comment
* use type_convert
* clean
* clean element wise op
[ROCm/composable_kernel commit: 56adf7e9cc]
This commit is contained in:
Chao Liu
@@ -40,6 +40,7 @@ set(PROFILER_SOURCE
|
||||
src/profile_grouped_gemm.cpp
|
||||
src/profile_conv_bwd_weight.cpp
|
||||
src/profile_batched_gemm_reduce.cpp
|
||||
src/profile_gemm_add_add_fastgelu.cpp
|
||||
)
|
||||
|
||||
add_executable(ckProfiler ${PROFILER_SOURCE})
|
||||
@@ -64,3 +65,4 @@ target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
|
||||
target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
|
||||
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance)
|
||||
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
|
||||
target_link_libraries(ckProfiler PRIVATE device_gemm_add_add_fastgelu_instance)
|
||||
|
||||
288
profiler/include/profile_gemm_add_add_fastgelu_impl.hpp
Normal file
288
profiler/include/profile_gemm_add_add_fastgelu_impl.hpp
Normal file
@@ -0,0 +1,288 @@
|
||||
#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 "reference_gemm.hpp"
|
||||
#include "device_gemm_multiple_d.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
|
||||
152
profiler/src/profile_gemm_add_add_fastgelu.cpp
Normal file
152
profiler/src/profile_gemm_add_add_fastgelu.cpp
Normal file
@@ -0,0 +1,152 @@
|
||||
#include <iostream>
|
||||
#include <numeric>
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
#include <stdlib.h>
|
||||
|
||||
#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;
|
||||
}
|
||||
}
|
||||
@@ -22,9 +22,39 @@ int profile_convnd_bwd_data(int, char*[], int);
|
||||
int profile_reduce(int, char*[]);
|
||||
int profile_conv_bwd_weight(int, char*[]);
|
||||
int profile_batched_gemm_reduce(int, char*[]);
|
||||
int profile_gemm_add_add_fastgelu(int, char*[]);
|
||||
|
||||
static void print_helper_message()
|
||||
{
|
||||
// clang-format off
|
||||
printf("arg1: tensor operation (gemm: GEMM\n"
|
||||
" gemm_bias_2d: GEMM+Bias(2D)\n"
|
||||
" gemm_bias_relu: GEMM+Bias+ReLU\n"
|
||||
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
|
||||
" gemm_reduce: GEMM+Reduce\n"
|
||||
" grouped_gemm: Grouped GEMM\n"
|
||||
" conv_fwd: ForwardConvolution\n"
|
||||
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
|
||||
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
|
||||
" conv_fwd_bias_relu_atomic_add: ForwardConvolution+Bias+ReLU+AtomicAdd\n"
|
||||
" conv1d_bwd_data: BackwardConvolution data 1 dim\n"
|
||||
" conv2d_bwd_data: BackwardConvolution data 2 dim\n"
|
||||
" conv3d_bwd_data: BackwardConvolution data 3 dim\n"
|
||||
" reduce: Reduce\n"
|
||||
" conv2d_bwd_weight: Backward Weight Convolution 2d\n"
|
||||
" gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n");
|
||||
// clang-format on
|
||||
}
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
if(argc == 1)
|
||||
{
|
||||
print_helper_message();
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
if(strcmp(argv[1], "gemm") == 0)
|
||||
{
|
||||
return profile_gemm(argc, argv);
|
||||
@@ -97,25 +127,14 @@ int main(int argc, char* argv[])
|
||||
{
|
||||
return profile_conv_bwd_weight(argc, argv);
|
||||
}
|
||||
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
|
||||
{
|
||||
return profile_gemm_add_add_fastgelu(argc, argv);
|
||||
}
|
||||
else
|
||||
{
|
||||
// clang-format off
|
||||
printf("arg1: tensor operation (gemm: GEMM\n"
|
||||
" gemm_bias_2d: GEMM+Bias(2D)\n"
|
||||
" gemm_bias_relu: GEMM+Bias+ReLU\n"
|
||||
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
|
||||
" gemm_reduce: GEMM+Reduce\n"
|
||||
" grouped_gemm: Grouped GEMM\n"
|
||||
" conv_fwd: ForwardConvolution\n"
|
||||
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
|
||||
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
|
||||
" conv_fwd_bias_relu_atomic_add: ForwardConvolution+Bias+ReLU+AtomicAdd\n"
|
||||
" conv1d_bwd_data: BackwardConvolution data 1 dim\n"
|
||||
" conv2d_bwd_data: BackwardConvolution data 2 dim\n"
|
||||
" conv3d_bwd_data: BackwardConvolution data 3 dim\n"
|
||||
" reduce: Reduce\n"
|
||||
" conv2d_bwd_weight: Backward Weight Convolution 2d\n");
|
||||
// clang-format on
|
||||
print_helper_message();
|
||||
|
||||
return 0;
|
||||
}
|
||||
return 0;
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user