Skip to content

GitLab

Unverified Commit 9e4429f9 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Gemm+Bilinear (#316) 

* refactor

* update example

* update example

* gemm bilinear

* clean

* update
parent 8e374781
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 515 additions and 1205 deletions
profiler/CMakeLists.txt
View file @ 9e4429f9
...@@ -7,21 +7,19 @@ set(PROFILER_SOURCE ...@@ -7,21 +7,19 @@ set(PROFILER_SOURCE
src/profiler.cpp src/profiler.cpp
src/profile_gemm.cpp src/profile_gemm.cpp
src/profile_gemm_splitk.cpp src/profile_gemm_splitk.cpp
src/profile_gemm_bias_2d.cpp src/profile_gemm_bilinear.cpp
src/profile_gemm_bias_relu.cpp
src/profile_gemm_bias_relu_add.cpp
src/profile_gemm_reduce.cpp
src/profile_gemm_bias_add_reduce.cpp src/profile_gemm_bias_add_reduce.cpp
src/profile_gemm_add_add_fastgelu.cpp
src/profile_gemm_reduce.cpp
src/profile_batched_gemm.cpp src/profile_batched_gemm.cpp
src/profile_batched_gemm_reduce.cpp
src/profile_grouped_gemm.cpp
src/profile_conv_fwd_bias_relu.cpp src/profile_conv_fwd_bias_relu.cpp
src/profile_conv_fwd_bias_relu_add.cpp src/profile_conv_fwd_bias_relu_add.cpp
src/profile_convnd_fwd.cpp src/profile_convnd_fwd.cpp
src/profile_convnd_bwd_data.cpp src/profile_convnd_bwd_data.cpp
src/profile_reduce.cpp
src/profile_grouped_gemm.cpp
src/profile_conv_bwd_weight.cpp src/profile_conv_bwd_weight.cpp
src/profile_batched_gemm_reduce.cpp src/profile_reduce.cpp
src/profile_gemm_add_add_fastgelu.cpp
src/profile_normalization.cpp src/profile_normalization.cpp
) )
...@@ -31,12 +29,10 @@ target_link_libraries(ckProfiler PRIVATE host_tensor) ...@@ -31,12 +29,10 @@ target_link_libraries(ckProfiler PRIVATE host_tensor)
target_link_libraries(ckProfiler PRIVATE conv_util) target_link_libraries(ckProfiler PRIVATE conv_util)
target_link_libraries(ckProfiler PRIVATE device_gemm_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_splitk_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_splitk_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_bilinear_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_add_add_fastgelu_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_add_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_bias_add_reduce_instance) target_link_libraries(ckProfiler PRIVATE device_gemm_bias_add_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_add_add_fastgelu_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_instance) target_link_libraries(ckProfiler PRIVATE device_batched_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance) target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance) target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
......
profiler/include/profile_batched_gemm_impl.hpp
View file @ 9e4429f9
...@@ -159,10 +159,10 @@ bool profile_batched_gemm_impl(int do_verification, ...@@ -159,10 +159,10 @@ bool profile_batched_gemm_impl(int do_verification,
BatchStrideA, BatchStrideA,
BatchStrideB, BatchStrideB,
BatchStrideC, BatchStrideC,
BatchCount,
ck::tensor_operation::element_wise::PassThrough{}, ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{}, ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{}, ck::tensor_operation::element_wise::PassThrough{});
BatchCount);
auto invoker_ptr = op_ptr->MakeInvokerPointer(); auto invoker_ptr = op_ptr->MakeInvokerPointer();
......
profiler/include/profile_gemm_bias_2d_impl.hpp deleted 100644 → 0
View file @ 8e374781
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#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 {
namespace device {
namespace instance {
using DeviceGemmAlphaBetaPtr = ck::tensor_operation::device::DeviceGemmBiasPtr<
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AlphaBetaAdd>;
void add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_km_kn_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f32_f32_f32_mk_nk_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
void add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmAlphaBetaPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename C0DataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout>
void profile_gemm_bias_2d_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,
float alpha,
float beta)
{
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<C0DataType> c0_m_n(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c0_m_n: " << c0_m_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
std::size_t num_thread = 1;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
c0_m_n.GenerateTensorValue(GeneratorTensor_2<C0DataType>{-5, 5}, num_thread);
break;
default:
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);
c0_m_n.GenerateTensorValue(GeneratorTensor_3<C0DataType>{-0.5, 0.5}, num_thread);
}
// set zero to c_device_buf
c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, 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::AlphaBetaAdd;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{alpha, beta};
if(do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBias2D<ADataType,
BDataType,
C0DataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c0_m_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem c0_device_buf(sizeof(C0DataType) * c0_m_n.mDesc.GetElementSpace());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
c0_device_buf.ToDevice(c0_m_n.mData.data());
c_device_buf.ToDevice(c_m_n_device_result.mData.data());
// add device GEMM instances
std::vector<ck::tensor_operation::device::instance::DeviceGemmAlphaBetaPtr> 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::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_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::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_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::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_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::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
}
else if constexpr(is_same<ADataType, float>::value && is_same<BDataType, float>::value &&
is_same<CDataType, float>::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::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f32_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::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f32_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::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f32_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::instance::
add_device_gemm_xdl_c_shuffle_bias_2d_f32_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<C0DataType*>(c0_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
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 * M + sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm_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());
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_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 << "c0 : ", c0_m_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;
}
}
}
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 deleted 100644 → 0
View file @ 8e374781
This diff is collapsed.
profiler/src/profile_batched_gemm.cpp
View file @ 9e4429f9
...@@ -27,8 +27,9 @@ enum struct GemmDataType ...@@ -27,8 +27,9 @@ enum struct GemmDataType
int profile_batched_gemm(int argc, char* argv[]) int profile_batched_gemm(int argc, char* argv[])
{ {
if(argc != 15) if(argc != 18)
{ {
// clang-format off
printf("arg1: tensor operation (batched_gemm: Batched GEMM)\n"); printf("arg1: tensor operation (batched_gemm: Batched GEMM)\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)\n");
printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n"); printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
...@@ -39,7 +40,8 @@ int profile_batched_gemm(int argc, char* argv[]) ...@@ -39,7 +40,8 @@ int profile_batched_gemm(int argc, char* argv[])
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\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("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n"); printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n"); printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n");
// clang-format on
exit(1); exit(1);
} }
...@@ -58,7 +60,11 @@ int profile_batched_gemm(int argc, char* argv[]) ...@@ -58,7 +60,11 @@ int profile_batched_gemm(int argc, char* argv[])
const int StrideB = std::stoi(argv[12]); const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]); const int StrideC = std::stoi(argv[13]);
const int BatchCount = std::stoi(argv[14]); const int BatchStrideA = std::stoi(argv[14]);
const int BatchStrideB = std::stoi(argv[15]);
const int BatchStrideC = std::stoi(argv[16]);
const int BatchCount = std::stoi(argv[17]);
using F32 = float; using F32 = float;
using F16 = ck::half_t; using F16 = ck::half_t;
...@@ -90,9 +96,13 @@ int profile_batched_gemm(int argc, char* argv[]) ...@@ -90,9 +96,13 @@ int profile_batched_gemm(int argc, char* argv[])
const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB; const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC; const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
const int BatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_; const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
const int BatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_; const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
const int BatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_; const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
bool pass = ck::profiler:: bool pass = ck::profiler::
profile_batched_gemm_impl<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>( profile_batched_gemm_impl<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>(
...@@ -103,9 +113,9 @@ int profile_batched_gemm(int argc, char* argv[]) ...@@ -103,9 +113,9 @@ int profile_batched_gemm(int argc, char* argv[])
M, M,
N, N,
K, K,
BatchStrideA, BatchStrideA_,
BatchStrideB, BatchStrideB_,
BatchStrideC, BatchStrideC_,
StrideA_, StrideA_,
StrideB_, StrideB_,
StrideC_, StrideC_,
......
profiler/src/profile_gemm_add_add_fastgelu.cpp
View file @ 9e4429f9
...@@ -29,7 +29,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[]) ...@@ -29,7 +29,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
if(argc != 16) if(argc != 16)
{ {
// clang-format off // clang-format off
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+GeLU)\n"); printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU)\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)\n");
printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\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(" 1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
...@@ -39,7 +39,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[]) ...@@ -39,7 +39,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\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("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (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"); printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
// clang-format on // clang-format on
exit(1); exit(1);
} }
......
profiler/src/profile_gemm_bias_2d.cpp deleted 100644 → 0
View file @ 8e374781
This diff is collapsed.
profiler/src/profile_gemm_bias_relu.cpp deleted 100644 → 0
View file @ 8e374781
This diff is collapsed.
profiler/src/profile_gemm_bias_relu_add.cpp deleted 100644 → 0
View file @ 8e374781
This diff is collapsed.
profiler/src/profile_gemm_bilinear.cpp 0 → 100644
View file @ 9e4429f9
This diff is collapsed.
profiler/src/profiler.cpp
View file @ 9e4429f9
...@@ -5,12 +5,10 @@ ...@@ -5,12 +5,10 @@
int profile_gemm(int, char*[]); int profile_gemm(int, char*[]);
int profile_gemm_splitk(int, char*[]); int profile_gemm_splitk(int, char*[]);
int profile_gemm_bias_2d(int, char*[]); int profile_gemm_bilinear(int, char*[]);
int profile_gemm_bias_relu(int, char*[]);
int profile_gemm_bias_relu_add(int, char*[]);
int profile_gemm_bias_add_reduce(int, char*[]);
int profile_gemm_add_add_fastgelu(int, char*[]); int profile_gemm_add_add_fastgelu(int, char*[]);
int profile_gemm_reduce(int, char*[]); int profile_gemm_reduce(int, char*[]);
int profile_gemm_bias_add_reduce(int, char*[]);
int profile_batched_gemm(int, char*[]); int profile_batched_gemm(int, char*[]);
int profile_batched_gemm_reduce(int, char*[]); int profile_batched_gemm_reduce(int, char*[]);
int profile_grouped_gemm(int, char*[]); int profile_grouped_gemm(int, char*[]);
...@@ -28,12 +26,12 @@ static void print_helper_message() ...@@ -28,12 +26,12 @@ static void print_helper_message()
// clang-format off // clang-format off
printf("arg1: tensor operation (gemm: GEMM\n" printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_splitk: Split-K GEMM\n" " gemm_splitk: Split-K GEMM\n"
" gemm_bias_2d: GEMM+Bias(2D)\n" " gemm_bilinear: GEMM+Bilinear\n"
" gemm_bias_relu: GEMM+Bias+ReLU\n"
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
" gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n" " gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n"
" gemm_reduce: GEMM+Reduce\n" " gemm_reduce: GEMM+Reduce\n"
" gemm_bias_add_reduce: GEMM+Bias+Add+Reduce\n"
" batched_gemm: Batched GEMM\n" " batched_gemm: Batched GEMM\n"
" batched_gemm_reduce: Batched GEMM+Reduce\n"
" grouped_gemm: Grouped GEMM\n" " grouped_gemm: Grouped GEMM\n"
" conv_fwd: ForwardConvolution\n" " conv_fwd: ForwardConvolution\n"
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n" " conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
...@@ -63,17 +61,13 @@ int main(int argc, char* argv[]) ...@@ -63,17 +61,13 @@ int main(int argc, char* argv[])
{ {
return profile_gemm_splitk(argc, argv); return profile_gemm_splitk(argc, argv);
} }
else if(strcmp(argv[1], "gemm_bias_2d") == 0) else if(strcmp(argv[1], "gemm_bilinear") == 0)
{
return profile_gemm_bias_2d(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_relu") == 0)
{ {
return profile_gemm_bias_relu(argc, argv); return profile_gemm_bilinear(argc, argv);
} }
else if(strcmp(argv[1], "gemm_bias_relu_add") == 0) else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
{ {
return profile_gemm_bias_relu_add(argc, argv); return profile_gemm_add_add_fastgelu(argc, argv);
} }
else if(strcmp(argv[1], "gemm_reduce") == 0) else if(strcmp(argv[1], "gemm_reduce") == 0)
{ {
...@@ -119,17 +113,13 @@ int main(int argc, char* argv[]) ...@@ -119,17 +113,13 @@ int main(int argc, char* argv[])
{ {
return profile_convnd_bwd_data(argc, argv, 3); return profile_convnd_bwd_data(argc, argv, 3);
} }
else if(strcmp(argv[1], "reduce") == 0)
{
return profile_reduce(argc, argv);
}
else if(strcmp(argv[1], "conv2d_bwd_weight") == 0) else if(strcmp(argv[1], "conv2d_bwd_weight") == 0)
{ {
return profile_conv_bwd_weight(argc, argv); return profile_conv_bwd_weight(argc, argv);
} }
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0) else if(strcmp(argv[1], "reduce") == 0)
{ {
return profile_gemm_add_add_fastgelu(argc, argv); return profile_reduce(argc, argv);
} }
else if(strcmp(argv[1], "batchnorm") == 0 || strcmp(argv[1], "layernorm") == 0 || else if(strcmp(argv[1], "batchnorm") == 0 || strcmp(argv[1], "layernorm") == 0 ||
strcmp(argv[1], "softmax") == 0) strcmp(argv[1], "softmax") == 0)
......
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