Skip to content

GitLab

Unverified Commit 29dcb956 authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge pull request #33 from ROCm/lwpck-1292 

Merge from the public repo.
parents 29deceb6 cbcc844e
Hide whitespace changes
Inline Side-by-side
Showing with 1125 additions and 320 deletions
profiler/include/profiler/profile_transpose_impl.hpp
View file @ 29dcb956
......@@ -25,7 +25,7 @@ namespace ck {
namespace profiler {
template <typename HostTensorA, typename HostTensorB, typename Functor>
void host_elementwise4D(HostTensorB& B_nchwd, const HostTensorA& A_ncdhw, Functor functor)
void host_elementwise4D(HostTensorB& B_ndhwc, const HostTensorA& A_ncdhw, Functor functor)
{
for(std::size_t n = 0; n < A_ncdhw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_ncdhw.mDesc.GetLengths()[1]; ++c)
......@@ -34,7 +34,7 @@ void host_elementwise4D(HostTensorB& B_nchwd, const HostTensorA& A_ncdhw, Functo
for(std::size_t w = 0; w < A_ncdhw.mDesc.GetLengths()[4]; ++w)
{
auto a_val = A_ncdhw(n, c, d, h, w);
functor(B_nchwd(n, c, h, w, d), a_val);
functor(B_ndhwc(n, d, h, w, c), a_val);
}
}
......@@ -77,8 +77,6 @@ bool profile_transpose_impl(int do_verification,
using ElementOp = ck::tensor_operation::element_wise::PassThrough;
// const auto element_op = ElementOp{};
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
......@@ -118,6 +116,7 @@ bool profile_transpose_impl(int do_verification,
// re-init C to zero before profiling next kernel
b_device_buf.SetZero();
// run for verification
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
if(do_verification)
......@@ -136,6 +135,7 @@ bool profile_transpose_impl(int do_verification,
std::string op_name = op_ptr->GetTypeString();
// run for timing purposes
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
......@@ -153,10 +153,6 @@ bool profile_transpose_impl(int do_verification,
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
// pass = pass & ck::utils::check_err(b_device_result, b_host_result);
pass &= ck::utils::check_err(
b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
if(tflops > best_tflops)
{
best_op_name = op_name;
......
profiler/src/CMakeLists.txt
View file @ 29dcb956
......@@ -16,7 +16,11 @@ set(PROFILER_SOURCES
profile_grouped_conv_fwd.cpp
profile_grouped_conv_bwd_weight.cpp
profile_reduce.cpp
profile_groupnorm_bwd_data.cpp
profile_groupnorm_fwd.cpp
profile_layernorm_bwd_data.cpp
profile_layernorm_bwd_gamma_beta.cpp
profile_groupnorm_bwd_gamma_beta.cpp
profile_layernorm_fwd.cpp
profile_max_pool3d_fwd.cpp
profile_avg_pool3d_bwd.cpp
......@@ -27,6 +31,7 @@ set(PROFILER_SOURCES
profile_batchnorm_infer.cpp
profile_grouped_conv_bwd_data.cpp
profile_conv_tensor_rearrange.cpp
profile_transpose.cpp
)
if(DL_KERNELS)
......@@ -38,7 +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)
......@@ -56,7 +64,7 @@ set(PROFILER_EXECUTABLE ckProfiler)
add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES})
target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
......@@ -78,6 +86,8 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_w
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_bwd_gamma_beta_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
......@@ -88,6 +98,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_d
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
......@@ -101,7 +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.cpp
View file @ 29dcb956
......@@ -42,12 +42,15 @@ static void print_helper_msg()
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0: no, 1: yes)\n"
<< "arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n"
<< "optional:\n"
<< "arg14: number of warm-up cycles (default 1)\n"
<< "arg15: number of iterations (default 10)\n"
<< std::endl;
}
int profile_gemm(int argc, char* argv[])
{
if(argc != 14)
if(argc != 14 && argc != 16)
{
print_helper_msg();
exit(1);
......@@ -68,6 +71,13 @@ int profile_gemm(int argc, char* argv[])
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
int n_warmup = 1;
int n_iter = 10;
if(argc == 16)
{
n_warmup = std::stoi(argv[14]);
n_iter = std::stoi(argv[15]);
}
using F32 = float;
using F16 = ck::half_t;
#ifdef CK_ENABLE_BF16
......@@ -120,13 +130,21 @@ int profile_gemm(int argc, char* argv[])
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideC < 0) ? DefaultStrideC : StrideC);
(StrideC < 0) ? DefaultStrideC : StrideC,
n_warmup,
n_iter);
return pass ? 0 : 1;
};
if(false)
;
if(data_type != GemmDataType::F32_F32_F32 && data_type != GemmDataType::F16_F16_F16 &&
data_type != GemmDataType::BF16_BF16_BF16 && data_type != GemmDataType::INT8_INT8_INT8 &&
data_type != GemmDataType::F8_F8_F8)
{
// dummy clause before the else clauses for different data types
std::cout << "Gemm: this data_type is not implemented" << std::endl;
return 1;
}
#ifdef CK_ENABLE_FP32
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
......@@ -219,7 +237,7 @@ int profile_gemm(int argc, char* argv[])
#endif
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
std::cout << "Gemm: this data_type & layout is not implemented" << std::endl;
return 1;
}
......
profiler/src/profile_gemm_add.cpp 0 → 100644
View file @ 29dcb956
// 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 @ 29dcb956
// 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 @ 29dcb956
// 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 @ 29dcb956
// 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);
profiler/src/profile_gemm_splitk.cpp
View file @ 29dcb956
......@@ -33,7 +33,7 @@ enum struct GemmDataType
int profile_gemm_splitk(int argc, char* argv[])
{
if(argc != 15)
if(argc != 15 && argc != 17)
{
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: f16, "
......@@ -48,6 +48,9 @@ int profile_gemm_splitk(int argc, char* argv[])
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: split k into mulitiple batch\n");
printf("optional:\n");
printf("arg15: number of warm-up cycles (default 1)\n");
printf("arg16: number of iterations (default 10)\n");
exit(1);
}
......@@ -67,6 +70,14 @@ int profile_gemm_splitk(int argc, char* argv[])
const int StrideC = std::stoi(argv[13]);
const int KBatch = std::stoi(argv[14]);
int n_warmup = 1;
int n_iter = 10;
if(argc == 17)
{
n_warmup = std::stoi(argv[15]);
n_iter = std::stoi(argv[16]);
}
using F32 = float;
using F16 = ck::half_t;
#if defined CK_ENABLE_FP8
......@@ -117,7 +128,9 @@ int profile_gemm_splitk(int argc, char* argv[])
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideC < 0) ? DefaultStrideC : StrideC,
KBatch);
KBatch,
n_warmup,
n_iter);
return pass ? 0 : 1;
};
......
profiler/src/profile_grouped_gemm.cpp
View file @ 29dcb956
......@@ -69,7 +69,10 @@ int profile_grouped_gemm(int argc, char* argv[])
<< "arg7: time kernel (0=n0, 1=yes)\n"
<< "arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
"64,64 64,64 128,128)\n"
<< "arg15: kbatch value (default 4)\n"
<< "arg15: kbatch value (default 1)\n"
<< "optional:\n"
<< "arg16: number of warm-up cycles (default 1)\n"
<< "arg17: number of iterations (default 10)\n"
<< std::endl;
exit(1);
......@@ -90,6 +93,15 @@ int profile_grouped_gemm(int argc, char* argv[])
const auto StrideBs = argToIntArray(argv[12]);
const auto StrideCs = argToIntArray(argv[13]);
const int kbatch = argc == 15 ? std::stoi(argv[14]) : 1;
int n_warmup = 1;
int n_iter = 10;
if(argc == 17)
{
n_warmup = std::stoi(argv[16]);
n_iter = std::stoi(argv[17]);
}
#ifdef CK_ENABLE_FP16
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
......@@ -109,7 +121,9 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch);
kbatch,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
......@@ -129,7 +143,9 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch);
kbatch,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
......@@ -149,7 +165,9 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch);
kbatch,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
......@@ -169,7 +187,9 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch);
kbatch,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
......@@ -189,7 +209,9 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch);
kbatch,
n_warmup,
n_iter);
}
else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
......@@ -209,7 +231,9 @@ int profile_grouped_gemm(int argc, char* argv[])
StrideAs,
StrideBs,
StrideCs,
kbatch);
kbatch,
n_warmup,
n_iter);
}
else
{
......
profiler/src/profile_groupnorm_bwd_data.cpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_groupnorm_bwd_data_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
struct groupnormBwdDataArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{
if(std::string("--") + key == argv[i])
{
int pos = i;
while(++i < argc && argv[i][0] != '-') {}
int end = i;
for(int j = pos + 1; j < end; j++)
{
long_opts[key].push_back(std::stoi(argv[j]));
}
return true;
}
return false;
}
void operator()(int argc, char* argv[])
{
for(auto& kv : long_opts)
{
for(int i = 1; i < argc; i++)
{
if(parse_opt(argc, argv, kv.first, i))
break;
}
}
}
};
void print_help_groupnorm_bwd_data()
{
// eg: ckProfiler groupnorm_bwd_data 1 0 2 0 1 --length 1 16 16 32 40
std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
<< "arg2: verification (0: no; 1: yes)\n"
<< "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg4: print tensor value (0: no; 1: yes)\n"
<< "arg5: time kernel (0=no, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 1 16 16 32 40) \n"
<< std::endl;
}
int profile_groupnorm_bwd_data(int argc, char* argv[])
{
if(argc <= 2)
{
print_help_groupnorm_bwd_data();
return 0;
}
groupnormBwdDataArgParser arg_parser;
// short unnamed options
const ck::DataTypeEnum data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
// parse the long options
arg_parser(argc, argv);
const std::vector<index_t> length = arg_parser.long_opts["length"];
using F32 = float;
if(length.size() == 5)
{
if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_groupnorm_bwd_data_impl<F32, F32, F32, F32, F32, F32>(
do_verification, init_method, do_log, time_kernel, length);
}
else
{
throw std::runtime_error("not implemented yet");
}
}
else
{
throw std::runtime_error("length should be 5");
}
return 0;
}
REGISTER_PROFILER_OPERATION("groupnorm_bwd_data",
"Group Normalization",
profile_groupnorm_bwd_data);
profiler/src/profile_groupnorm_bwd_gamma_beta.cpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_groupnorm_bwd_gamma_beta_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
struct groupnormBwdGammaBetaArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{
if(std::string("--") + key == argv[i])
{
int pos = i;
while(++i < argc && argv[i][0] != '-') {}
int end = i;
for(int j = pos + 1; j < end; j++)
{
long_opts[key].push_back(std::stoi(argv[j]));
}
return true;
}
return false;
}
void operator()(int argc, char* argv[])
{
for(auto& kv : long_opts)
{
for(int i = 1; i < argc; i++)
{
if(parse_opt(argc, argv, kv.first, i))
break;
}
}
}
};
void print_help_groupnorm_bwd_gamma_beta()
{
// eg: ckProfiler groupnorm_bwd_gamma_beta 1 0 2 0 1 --length 1 16 16 32 40
std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
<< "arg2: verification (0: no; 1: yes)\n"
<< "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg4: print tensor value (0: no; 1: yes)\n"
<< "arg5: time kernel (0=no, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 1 16 16 32 40) \n"
<< std::endl;
}
int profile_groupnorm_bwd_gamma_beta(int argc, char* argv[])
{
if(argc <= 2)
{
print_help_groupnorm_bwd_gamma_beta();
return 0;
}
groupnormBwdGammaBetaArgParser arg_parser;
// short unnamed options
const ck::DataTypeEnum data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
// parse the long options
arg_parser(argc, argv);
const std::vector<index_t> length = arg_parser.long_opts["length"];
using F32 = float;
if(length.size() == 5)
{
if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_groupnorm_bwd_gamma_beta_impl<F32, F32, F32, F32, F32, F32>(
do_verification, init_method, do_log, time_kernel, length);
}
else
{
throw std::runtime_error("not implemented yet");
}
}
else
{
throw std::runtime_error("length should be 5");
}
return 0;
}
REGISTER_PROFILER_OPERATION("groupnorm_bwd_gamma_beta",
"Group Normalization",
profile_groupnorm_bwd_gamma_beta);
profiler/src/profile_groupnorm_fwd.cpp
View file @ 29dcb956
......@@ -98,7 +98,7 @@ int profile_groupnorm(int argc, char* argv[])
}
else if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16, F32, false>(
ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16, F16, false>(
do_verification, init_method, do_log, time_kernel, length);
}
else
......
profiler/src/profile_layernorm_bwd_data.cpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_layernorm_bwd_data_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
struct layernormBwdDataArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{
if(std::string("--") + key == argv[i])
{
int pos = i;
while(++i < argc && argv[i][0] != '-') {}
int end = i;
for(int j = pos + 1; j < end; j++)
{
long_opts[key].push_back(std::stoi(argv[j]));
}
return true;
}
return false;
}
void operator()(int argc, char* argv[])
{
for(auto& kv : long_opts)
{
for(int i = 1; i < argc; i++)
{
if(parse_opt(argc, argv, kv.first, i))
break;
}
}
}
};
void print_help_layernorm_bwd_data()
{
// eg: ckProfiler layernorm_bwd_data 0 0 2 0 1 --length 1502 4096
std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
<< "arg2: verification (0: no; 1: yes)\n"
<< "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg4: print tensor value (0: no; 1: yes)\n"
<< "arg5: time kernel (0=no, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 1024 1024) \n"
<< std::endl;
}
int profile_layernorm_bwd_data(int argc, char* argv[])
{
if(argc <= 2)
{
print_help_layernorm_bwd_data();
return 0;
}
layernormBwdDataArgParser arg_parser;
// short unnamed options
const ck::DataTypeEnum data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
// parse the long options
arg_parser(argc, argv);
const std::vector<index_t> length = arg_parser.long_opts["length"];
using F16 = ck::half_t;
using F32 = float;
if(length.size() == 2)
{
constexpr int rank = 2;
if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_layernorm_bwd_data_impl<F16, F16, F16, F16, F32, F16, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_layernorm_bwd_data_impl<F32, F32, F32, F32, F32, F32, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else
{
throw std::runtime_error("not implemented yet");
}
}
else
{
throw std::runtime_error("not implemented yet");
}
return 0;
}
REGISTER_PROFILER_OPERATION("layernorm_bwd_data",
"Layer Normalization",
profile_layernorm_bwd_data);
profiler/src/profile_layernorm_bwd_gamma_beta.cpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_layernorm_bwd_gamma_beta_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
struct layernormBwdGammaBetaArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
bool parse_opt(int argc, char* argv[], const std::string& key, int i)
{
if(std::string("--") + key == argv[i])
{
int pos = i;
while(++i < argc && argv[i][0] != '-') {}
int end = i;
for(int j = pos + 1; j < end; j++)
{
long_opts[key].push_back(std::stoi(argv[j]));
}
return true;
}
return false;
}
void operator()(int argc, char* argv[])
{
for(auto& kv : long_opts)
{
for(int i = 1; i < argc; i++)
{
if(parse_opt(argc, argv, kv.first, i))
break;
}
}
}
};
void print_help_layernorm_bwd_gamma_beta()
{
// eg: ckProfiler layernorm_bwd_gamma_beta 0 0 2 0 1 --length 1502 4096
std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
<< "arg2: verification (0: no; 1: yes)\n"
<< "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg4: print tensor value (0: no; 1: yes)\n"
<< "arg5: time kernel (0=no, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 1024 1024) \n"
<< std::endl;
}
int profile_layernorm_bwd_gamma_beta(int argc, char* argv[])
{
if(argc <= 2)
{
print_help_layernorm_bwd_gamma_beta();
return 0;
}
layernormBwdGammaBetaArgParser arg_parser;
// short unnamed options
const ck::DataTypeEnum data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
// parse the long options
arg_parser(argc, argv);
const std::vector<index_t> length = arg_parser.long_opts["length"];
using F16 = ck::half_t;
using F32 = float;
if(length.size() == 2)
{
constexpr int rank = 2;
if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_layernorm_bwd_gamma_beta_impl<F16, F16, F16, F32, F16, F16, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_layernorm_bwd_gamma_beta_impl<F32, F32, F32, F32, F32, F32, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else
{
throw std::runtime_error("not implemented yet");
}
}
else
{
throw std::runtime_error("not implemented yet");
}
return 0;
}
REGISTER_PROFILER_OPERATION("layernorm_bwd_gamma_beta",
"Layer Normalization",
profile_layernorm_bwd_gamma_beta);
profiler/src/profile_layernorm_fwd.cpp
View file @ 29dcb956
......@@ -104,7 +104,7 @@ int profile_layernorm(int argc, char* argv[])
if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F32, false, rank>(
ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, F16, false, rank>(
do_verification, init_method, do_log, time_kernel, length);
}
else if(data_type == ck::DataTypeEnum::Float)
......@@ -125,4 +125,4 @@ int profile_layernorm(int argc, char* argv[])
return 0;
}
REGISTER_PROFILER_OPERATION("layernorm", "Layer Normalization", profile_layernorm);
REGISTER_PROFILER_OPERATION("layernorm_fwd", "Layer Normalization", profile_layernorm);
profiler/src/profile_transpose.cpp 0 → 100644
View file @ 29dcb956
// 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_transpose_impl.hpp"
#include "profiler_operation_registry.hpp"
enum struct DataType
{
F32_F32_F32_F32_F32, // 0
F16_F16_F16_F16_F16, // 1
};
#define OP_NAME "transpose"
#define OP_DESC "Transpose"
struct TransposeArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {{"lengths", {}}};
bool parse_opt(const int argc, char* argv[], const std::string& key, int i)
{
if(std::string("--") + key == argv[i])
{
const int pos = i;
while(++i < argc && argv[i][0] != '-') {}
int end = i;
for(int j = pos + 1; j < end; j++)
{
long_opts[key].push_back(std::stoi(argv[j]));
}
return true;
}
return false;
}
void operator()(int argc, char* argv[])
{
for(auto& kv : long_opts)
{
for(int i = 1; i < argc; i++)
{
if(parse_opt(argc, argv, kv.first, i))
break;
}
}
}
};
static void print_helper_msg()
{
printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: verification (0: no; 1: yes)\n");
printf("arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg5: print tensor value (0: no; 1: yes)\n");
printf("arg6: time kernel (0=no, 1=yes)\n");
printf("arg7: --lengths: N, C, D, H, W\n");
}
int profile_transpose(int argc, char* argv[])
{
if(argc != 7)
{
print_helper_msg();
exit(1);
}
TransposeArgParser arg_parser;
const auto data_type = static_cast<DataType>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
arg_parser(argc, argv);
const std::vector<ck::index_t> lengths = arg_parser.long_opts["lengths"];
using F32 = float;
using F16 = ck::half_t;
auto profile = [&](auto a_type, auto b_type) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
constexpr ck::index_t NumDim = 5;
bool pass = ck::profiler::profile_transpose_impl<ADataType, BDataType, NumDim>(
do_verification, init_method, do_log, time_kernel, lengths);
return pass ? 0 : 1;
};
if(data_type == DataType::F32_F32_F32_F32_F32)
{
return profile(F32{}, F32{});
}
else if(data_type == DataType::F16_F16_F16_F16_F16)
{
return profile(F16{}, F16{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_transpose);
script/clang-format-overwrite.sh
View file @ 29dcb956
#find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' -o -iname '*.inc' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'
find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' -o -iname '*.inc' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'
git status --porcelain | awk '$1 != "D" && (match($2, "\\.cpp|hpp|inc")) {print $2}' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'
script/parse_perf_data.py deleted 100644 → 0
View file @ 29deceb6
#!/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 @ 29dcb956
......@@ -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 @ 29dcb956
#!/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
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