Skip to content

GitLab

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

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

parents e07b3d8e d3051d75
Hide whitespace changes
Inline Side-by-side
Showing with 574 additions and 539 deletions
profiler/src/profile_gemm_bias_add_reduce.cpp 0 → 100644
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_bias_add_reduce_impl.hpp"
int profile_gemm_bias_add_reduce(int argc, char* argv[])
{
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmReduceDataType
{
F32_F32_F32_F32_F32_F32_F32, // 0
F16_F16_F16_F16_F16_F32_F32, // 1
};
if(!(argc == 14 || argc == 15))
{
printf("arg1: tensor operation (gemm: GEMM+bias+add+Reduce)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[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=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, StrideC1\n");
exit(1);
}
const auto data_type = static_cast<GemmReduceDataType>(std::stoi(argv[2]));
const auto layout = static_cast<GemmMatrixLayout>(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 StrideC = std::stoi(argv[13]);
const int StrideC1 = std::stoi(argv[14]);
if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? M : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
(StrideC1 < 0) ? N : StrideC1);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 0;
}
profiler/src/profile_gemm_bias_relu.cpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "profiler/include/profile_gemm_bias_relu_impl.hpp"
#include "profile_gemm_bias_relu_impl.hpp"
enum struct GemmMatrixLayout enum struct GemmMatrixLayout
{ {
......
profiler/src/profile_gemm_bias_relu_add.cpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "profiler/include/profile_gemm_bias_relu_add_impl.hpp"
#include "profile_gemm_bias_relu_add_impl.hpp"
enum struct GemmMatrixLayout enum struct GemmMatrixLayout
{ {
......
profiler/src/profile_gemm_reduce.cpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "profiler/include/profile_gemm_reduce_impl.hpp"
#include "profile_gemm_reduce_impl.hpp"
int profile_gemm_reduce(int argc, char* argv[]) int profile_gemm_reduce(int argc, char* argv[])
{ {
......
profiler/src/profile_grouped_gemm.cpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <stdlib.h>
#include <half.hpp> #include "profiler/include/profile_grouped_gemm_impl.hpp"
#include "profile_grouped_gemm_impl.hpp"
enum struct GemmMatrixLayout enum struct GemmMatrixLayout
{ {
......
profiler/src/profile_reduce.cpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <fstream> #include <fstream>
#include <cstdlib> #include <cstdlib>
...@@ -6,11 +9,12 @@ ...@@ -6,11 +9,12 @@
#include <sstream> #include <sstream>
#include <getopt.h> #include <getopt.h>
#include "data_type_enum.hpp" #include "ck/utility/reduction_enums.hpp"
#include "reduction_enums.hpp"
#include "ck/library/host_tensor/host_common_util.hpp"
#include "host_common_util.hpp" #include "profiler/include/profile_reduce_impl.hpp"
#include "profile_reduce_impl.hpp" #include "profiler/include/data_type_enum.hpp"
using namespace std; using namespace std;
......
profiler/src/profiler.cpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
#include <initializer_list> #include <initializer_list>
#include <cstdlib> #include <cstdlib>
#include <cstring> #include <cstring>
#include "profile_convnd_fwd.hpp" #include "profiler/include/profile_convnd_fwd.hpp"
int profile_gemm(int, char*[]); int profile_gemm(int, char*[]);
int profile_gemm_bias_2d(int, char*[]); int profile_gemm_bias_2d(int, char*[]);
int profile_gemm_bias_relu(int, char*[]); int profile_gemm_bias_relu(int, char*[]);
int profile_gemm_bias_relu_add(int, char*[]); int profile_gemm_bias_relu_add(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_grouped_gemm(int, char*[]); int profile_grouped_gemm(int, char*[]);
int profile_conv_fwd(int, char*[]); int profile_conv_fwd(int, char*[]);
int profile_conv_fwd_bias_relu(int, char*[]); int profile_conv_fwd_bias_relu(int, char*[]);
int profile_conv_fwd_bias_relu_add(int, char*[]); int profile_conv_fwd_bias_relu_add(int, char*[]);
int profile_conv_fwd_bias_relu_atomic_add(int, char*[]);
int profile_convnd_bwd_data(int, char*[], int); int profile_convnd_bwd_data(int, char*[], int);
int profile_reduce(int, char*[]); int profile_reduce(int, char*[]);
int profile_conv_bwd_weight(int, char*[]); int profile_conv_bwd_weight(int, char*[]);
int profile_batched_gemm_reduce(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"
" 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[]) int main(int argc, char* argv[])
{ {
if(argc == 1)
{
print_helper_message();
return 0;
}
if(strcmp(argv[1], "gemm") == 0) if(strcmp(argv[1], "gemm") == 0)
{ {
return profile_gemm(argc, argv); return profile_gemm(argc, argv);
...@@ -44,6 +76,10 @@ int main(int argc, char* argv[]) ...@@ -44,6 +76,10 @@ int main(int argc, char* argv[])
{ {
return profile_gemm_reduce(argc, argv); return profile_gemm_reduce(argc, argv);
} }
else if(strcmp(argv[1], "gemm_bias_add_reduce") == 0)
{
return profile_gemm_bias_add_reduce(argc, argv);
}
else if(strcmp(argv[1], "batched_gemm") == 0) else if(strcmp(argv[1], "batched_gemm") == 0)
{ {
return profile_batched_gemm(argc, argv); return profile_batched_gemm(argc, argv);
...@@ -68,10 +104,6 @@ int main(int argc, char* argv[]) ...@@ -68,10 +104,6 @@ int main(int argc, char* argv[])
{ {
return profile_conv_fwd_bias_relu_add(argc, argv); return profile_conv_fwd_bias_relu_add(argc, argv);
} }
else if(strcmp(argv[1], "conv_fwd_bias_relu_atomic_add") == 0)
{
return profile_conv_fwd_bias_relu_atomic_add(argc, argv);
}
else if(strcmp(argv[1], "conv1d_bwd_data") == 0) else if(strcmp(argv[1], "conv1d_bwd_data") == 0)
{ {
return profile_convnd_bwd_data(argc, argv, 1); return profile_convnd_bwd_data(argc, argv, 1);
...@@ -92,25 +124,14 @@ int main(int argc, char* argv[]) ...@@ -92,25 +124,14 @@ int main(int argc, char* argv[])
{ {
return profile_conv_bwd_weight(argc, 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 else
{ {
// clang-format off print_helper_message();
printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_bias_2d: GEMM+Bias(2D)\n" return 0;
" 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
} }
return 0;
} }
script/parse_perf_data.py
View file @ 7a3b49e5
#!/usr/bin/env python3 #!/usr/bin/env python3
import os, io, argparse, datetime import os, io, argparse, datetime, re
import numpy as np import numpy as np
import sqlalchemy import sqlalchemy
from sqlalchemy.types import NVARCHAR, Float, Integer from sqlalchemy.types import NVARCHAR, Float, Integer
...@@ -45,66 +45,98 @@ def main(): ...@@ -45,66 +45,98 @@ def main():
StrideB=[] StrideB=[]
StrideC=[] StrideC=[]
#parse results, get the Tflops value for "Best Perf" kernels #parse results, get the Tflops value for "Best Perf" kernels
glue="" glue=""
for filename in args.files: for filename in args.files:
for line in open(filename): for line in open(filename):
if 'Branch name' in line: if 'Branch name' in line:
lst=line.split() lst=line.split()
branch_name=lst[2] branch_name=lst[2]
for filename in args.files: if 'On branch' in line:
for line in open(filename): lst=line.split()
if 'Best Perf' in line: branch_name=lst[2]
if 'Node name' in line:
lst=line.split() lst=line.split()
if len(lst)>=37: #the line is complete node_id=lst[2]
tests.append(glue.join(lst[5:30])) if 'GPU_arch' in line:
kernels.append(glue.join(lst[37:])) lst=line.split()
tflops.append(lst[33]) gpu_arch=lst[2]
dtype.append(lst[5]) if 'HIP version' in line:
alayout.append(lst[8]) lst=line.split()
blayout.append(lst[11]) hip_vers=lst[2]
M.append(lst[14]) if 'Compute Unit' in line:
N.append(lst[17]) lst=line.split()
K.append(lst[20]) compute_units=lst[2]
StrideA.append(lst[23]) if 'InstalledDir' in line:
StrideB.append(lst[26]) lst=line.split()
StrideC.append(lst[29]) rocm_vers=lst[1][lst[1].find('/opt/rocm-')+len('/opt/rocm-'):lst[1].rfind('/llvm/bin')]
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
print("Number of tests:",len(tests))
print("Branch name:",branch_name) print("Branch name:",branch_name)
#sorted_tests = sorted(tests) print("Node name:",node_id)
#print("sorted tests:",sorted_tests) print("GPU_arch:",gpu_arch)
sorted_tflops = [x for _,x in sorted(zip(tests,tflops))] print("Compute units:",compute_units)
#sorted_kernels = [x for _,x in sorted(zip(tests,kernels))] print("ROCM_version:",rocm_vers)
test_list=list(range(1,len(tests)+1)) 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_hostname = '127.0.0.1'
sql_username = os.environ["dbuser"] sql_username = os.environ["dbuser"]
print("sql_username=",sql_username)
sql_password = os.environ["dbpassword"] sql_password = os.environ["dbpassword"]
sql_main_database = 'miopen_perf' sql_main_database = 'miopen_perf'
sql_port = 3306 sql_port = 3306
ssh_host = os.environ["dbsship"] ssh_host = os.environ["dbsship"]
print("ssh_host=",ssh_host)
ssh_user = os.environ["dbsshuser"] ssh_user = os.environ["dbsshuser"]
print("ssh_user=",ssh_user)
ssh_port = int(os.environ["dbsshport"]) ssh_port = int(os.environ["dbsshport"])
ssh_pass = os.environ["dbsshpassword"] ssh_pass = os.environ["dbsshpassword"]
...@@ -118,75 +150,140 @@ def main(): ...@@ -118,75 +150,140 @@ def main():
format(sql_username, sql_password, sql_hostname, tunnel.local_bind_port, sql_main_database)) format(sql_username, sql_password, sql_hostname, tunnel.local_bind_port, sql_main_database))
conn = sqlEngine.connect() conn = sqlEngine.connect()
#write the ck_gemm_test_params table #save gemm performance tests:
#only needed once the test set changes if 'gemm' in filename:
'''
sorted_dtypes = [x for _,x in sorted(zip(tests,dtype))] #write the ck_gemm_test_params table
sorted_alayout = [x for _,x in sorted(zip(tests,alayout))] #only needed once the test set changes
sorted_blayout = [x for _,x in sorted(zip(tests,blayout))] '''
sorted_M = [x for _,x in sorted(zip(tests,M))] sorted_dtypes = [x for _,x in sorted(zip(tests,dtype))]
sorted_N = [x for _,x in sorted(zip(tests,N))] sorted_alayout = [x for _,x in sorted(zip(tests,alayout))]
sorted_K = [x for _,x in sorted(zip(tests,K))] sorted_blayout = [x for _,x in sorted(zip(tests,blayout))]
sorted_StrideA = [x for _,x in sorted(zip(tests,StrideA))] sorted_M = [x for _,x in sorted(zip(tests,M))]
sorted_StrideB = [x for _,x in sorted(zip(tests,StrideB))] sorted_N = [x for _,x in sorted(zip(tests,N))]
sorted_StrideC = [x for _,x in sorted(zip(tests,StrideC))] sorted_K = [x for _,x in sorted(zip(tests,K))]
ck_gemm_params=[test_list,sorted_dtypes,sorted_alayout,sorted_blayout, sorted_StrideA = [x for _,x in sorted(zip(tests,StrideA))]
sorted_M,sorted_N,sorted_K,sorted_StrideA,sorted_StrideB, sorted_StrideB = [x for _,x in sorted(zip(tests,StrideB))]
sorted_StrideC] sorted_StrideC = [x for _,x in sorted(zip(tests,StrideC))]
df=pd.DataFrame(np.transpose(ck_gemm_params),columns=['Test_number','Data_type', ck_gemm_params=[test_list,sorted_dtypes,sorted_alayout,sorted_blayout,
'Alayout','BLayout','M','N','K', 'StrideA','StrideB','StrideC']) sorted_M,sorted_N,sorted_K,sorted_StrideA,sorted_StrideB,
print(df) sorted_StrideC]
df=pd.DataFrame(np.transpose(ck_gemm_params),columns=['Test_number','Data_type',
dtypes = { 'Alayout','BLayout','M','N','K', 'StrideA','StrideB','StrideC'])
'Test_number': Integer(), print(df)
'Data_type': NVARCHAR(length=5),
'Alayout': NVARCHAR(length=12), dtypes = {
'Blayout': NVARCHAR(length=12), 'Test_number': Integer(),
'M': Integer(), 'Data_type': NVARCHAR(length=5),
'N': Integer(), 'Alayout': NVARCHAR(length=12),
'K': Integer(), 'Blayout': NVARCHAR(length=12),
'StrideA': Integer(), 'M': Integer(),
'StrideB': Integer(), 'N': Integer(),
'StrideC': Integer() 'K': Integer(),
} 'StrideA': Integer(),
df.to_sql("ck_gemm_test_params",conn,if_exists='replace',index=False, dtype=dtypes) 'StrideB': Integer(),
''' 'StrideC': Integer()
}
#read baseline results for the latest develop branch df.to_sql("ck_gemm_test_params",conn,if_exists='replace',index=False, dtype=dtypes)
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)
#read baseline results for the latest develop branch
#write new results to the db query = '''SELECT * from ck_gemm_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_gemm_tflops where Branch_ID='develop' );'''
testlist=[] tflops_base = pd.read_sql_query(query, conn)
for i in range(1,len(tests)+1):
testlist.append("Test%i"%i) #write new results to the db
ck_gemm_tflops=[str(branch_name),str(datetime.datetime.now())] testlist=[]
flops=pd.DataFrame(data=[ck_gemm_tflops],columns=['Branch_ID','Datetime']) for i in range(1,len(tests)+1):
df_add=pd.DataFrame(data=[sorted_tflops],columns=testlist) testlist.append("Test%i"%i)
flops=pd.concat([flops,df_add],axis=1) ck_gemm_tflops=[str(branch_name),str(node_id),str(gpu_arch),compute_units,str(rocm_vers),str(hip_vers),str(datetime.datetime.now())]
print("new tflops results:",flops) flops=pd.DataFrame(data=[ck_gemm_tflops],columns=['Branch_ID','Node_ID','GPU_arch','Compute Units','ROCM_version','HIP_version','Datetime'])
flops.to_sql("ck_gemm_tflops",conn,if_exists='append',index=False) 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() conn.close()
#compare the results to the baseline #compare the results to the baseline if baseline exists
regression=0 regression=0
base=tflops_base[testlist].to_numpy(dtype='float') if 'gemm' in filename:
base_list=base[0] if not tflops_base.empty:
ave_perf=0 base=tflops_base[testlist].to_numpy(dtype='float')
for i in range(len(base_list)): base_list=base[0]
# success criterion: ave_perf=0
if base_list[i]>1.01*float(sorted_tflops[i]): for i in range(len(base_list)):
print("test # ",i,"shows regression by {:.3f}%".format( # success criterion:
(float(sorted_tflops[i])-base_list[i])/base_list[i]*100)) if base_list[i]>1.01*float(sorted_tflops[i]):
regression=1 print("test # ",i,"shows regression by {:.3f}%".format(
ave_perf=ave_perf+float(sorted_tflops[i])/base_list[i] (float(sorted_tflops[i])-base_list[i])/base_list[i]*100))
if regression==0: regression=1
print("no regressions found") ave_perf=ave_perf+float(sorted_tflops[i])/base_list[i]
ave_perf=ave_perf/len(base_list) if regression==0:
print("average performance relative to baseline:",ave_perf) 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 0 if performance criteria met, otherwise return 1
return regression return regression
if __name__ == '__main__': if __name__ == '__main__':
......
script/profile_conv.sh
View file @ 7a3b49e5
...@@ -3,9 +3,9 @@ ...@@ -3,9 +3,9 @@
## GPU visibility ## GPU visibility
export HIP_VISIBLE_DEVICES=0 export HIP_VISIBLE_DEVICES=0
make -j ckProfiler # make -j ckProfiler
DRIVER="./profiler/ckProfiler" DRIVER="../build/bin/ckProfiler"
OP=$1 OP=$1
DATATYPE=$2 DATATYPE=$2
...@@ -26,7 +26,7 @@ REPEAT=$9 ...@@ -26,7 +26,7 @@ REPEAT=$9
N=${10} N=${10}
# Resnet50 from Bing # Resnet50 (no duplicated layer)
######## op datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads ######## op datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0 #$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 1024 1 1 14 14 1 1 1 1 0 0 0 0 #$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 1024 1 1 14 14 1 1 1 1 0 0 0 0
...@@ -47,60 +47,60 @@ REPEAT=$9 ...@@ -47,60 +47,60 @@ REPEAT=$9
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0 #$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 1 1 56 56 1 1 1 1 0 0 0 0 #$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 1 1 56 56 1 1 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1 #$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 8 7 7 224 224 2 2 1 1 3 3 3 3 #$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 3 7 7 224 224 2 2 1 1 3 3 3 3
# Resnet50 from Bing # Resnet50 fusion
#################### op____________________ datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads ####### op_________________ datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C_ Y X Hi_ Wi__ Strides Dilations LeftPads RightPads
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 3 7 7 224 224 2 2 1 1 3 3 3 3 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 3 7 7 224 224 2 2 1 1 3 3 3 3
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 1 1 56 56 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 256 1 1 56 56 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 256 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 256 1 1 56 56 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 256 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 256 1 1 56 56 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 256 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 56 56 2 2 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 56 56 2 2 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 512 1 1 28 28 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 512 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 28 28 2 2 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 28 28 2 2 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 1024 1 1 14 14 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 14 14 2 2 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 14 14 2 2 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 2048 1 1 7 7 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 2048 1 1 7 7 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 7 7 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 7 7 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 2048 1 1 7 7 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 2048 1 1 7 7 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 7 7 1 1 1 1 1 1 1 1 $DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 7 7 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0 $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
# Resnet50 # Resnet50
......
script/run_performance_tests.sh 0 → 100755
View file @ 7a3b49e5
#!/bin/bash
#
# in order to run this script you'd first need to build the ckProfiler executable in ../build/bin/
# and make sure the following python packages are installed in your environment:
pip3 install --upgrade pip
pip3 install sqlalchemy pymysql pandas sshtunnel
# you would also need to set up some environment variables in order to
# post your new test results to the database and compare them to the baseline
# please contact Illia.Silin@amd.com for more details
#
export gemm_log="perf_gemm.log"
rm -f $gemm_log
git status | grep -e 'On branch' > ${gemm_log}
echo -n 'Node name: ' >>${gemm_log}; hostname >> ${gemm_log}
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> ${gemm_log}; rocminfo | grep "Name:" | grep "gfx" >> ${gemm_log}
rocminfo | grep "Compute Unit:" >> ${gemm_log}
hipcc --version | grep -e 'HIP version' >> ${gemm_log}
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${gemm_log}
./profile_gemm.sh gemm 0 0 0 1 0 5 | tee -a ${gemm_log}
./profile_gemm.sh gemm 1 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 0 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 0 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 1 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 0 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 2 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 0 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 1 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 2 3 0 1 0 5 | tee -a $gemm_log
./profile_gemm.sh gemm 3 3 0 1 0 5 | tee -a $gemm_log
python3 parse_perf_data.py ${gemm_log}
#run resnet50 test
export resnet_log="perf_resnet50.log"
rm -f $resnet_log
git status | grep -e 'On branch' > ${resnet_log}
echo -n 'Node name: '>>${resnet_log}; hostname >>${resnet_log}
#get GPU_arch and number of compute units from rocminfo
echo -n "GPU_arch: " >> ${resnet_log}; rocminfo | grep "Name:" | grep "gfx" >> ${resnet_log}
rocminfo | grep "Compute Unit:" >> ${resnet_log}
hipcc --version | grep -e 'HIP version' >> ${resnet_log}
/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> ${resnet_log}
#first run tests with N=256
./profile_conv.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 256 | tee -a ${resnet_log}
#then run with N=4
./profile_conv.sh conv_fwd_bias_relu 1 1 1 1 0 2 0 1 4 | tee -a ${resnet_log}
#the script will put the results from N=256 and N=4 runs into separate tables
python3 parse_perf_data.py ${resnet_log}
test/CMakeLists.txt
View file @ 7a3b49e5
include_directories(BEFORE include_directories(BEFORE
${PROJECT_SOURCE_DIR}/ ${PROJECT_SOURCE_DIR}/
${PROJECT_SOURCE_DIR}/include/ck
${PROJECT_SOURCE_DIR}/include/ck/utility
${PROJECT_SOURCE_DIR}/include/ck/host_utility
${PROJECT_SOURCE_DIR}/include/ck/tensor_description
${PROJECT_SOURCE_DIR}/include/ck/tensor
${PROJECT_SOURCE_DIR}/include/ck/problem_transform
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/device
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/grid
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/block
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/warp
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/thread
${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/element
${PROJECT_SOURCE_DIR}/library/include/ck/library/host_tensor
${PROJECT_SOURCE_DIR}/library/include/ck/library/tensor_operation_instance
${PROJECT_SOURCE_DIR}/library/include/ck/library/tensor_operation_instance/gpu/reduce
${PROJECT_SOURCE_DIR}/library/include/ck/library/reference_tensor_operation/cpu
${PROJECT_SOURCE_DIR}/library/include/ck/library/reference_tensor_operation/gpu
${PROJECT_SOURCE_DIR}/library/include/ck/library/utility
${PROJECT_SOURCE_DIR}/test/include
${PROJECT_SOURCE_DIR}/profiler/include
${PROJECT_SOURCE_DIR}/external/include/half
) )
include(googletest) include(googletest)
...@@ -65,4 +44,4 @@ add_subdirectory(reduce) ...@@ -65,4 +44,4 @@ add_subdirectory(reduce)
add_subdirectory(conv2d_bwd_weight) add_subdirectory(conv2d_bwd_weight)
add_subdirectory(convnd_bwd_data) add_subdirectory(convnd_bwd_data)
add_subdirectory(block_to_ctile_map) add_subdirectory(block_to_ctile_map)
# DONOT add client_app, that is tested via CI independently add_subdirectory(softmax)
test/batched_gemm/batched_gemm_fp16.cpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include "profile_batched_gemm_impl.hpp" #include "profiler/include/profile_batched_gemm_impl.hpp"
namespace { namespace {
using ADataType = ck::half_t; using ADataType = ck::half_t;
......
test/batched_gemm/batched_gemm_util.hpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef BATCHED_GEMM_UTILS_HPP #ifndef BATCHED_GEMM_UTILS_HPP
#define BATCHED_GEMM_UTILS_HPP #define BATCHED_GEMM_UTILS_HPP
......
test/batched_gemm_reduce/CMakeLists.txt
View file @ 7a3b49e5
include_directories(BEFORE
${PROJECT_SOURCE_DIR}/profiler/include
${PROJECT_SOURCE_DIR}/test/include
${PROJECT_SOURCE_DIR}/external/include/half
)
add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp) add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE host_tensor) target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE host_tensor)
target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE device_batched_gemm_reduce_instance) target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE device_batched_gemm_reduce_instance)
test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream> #include <iostream>
#include "profile_batched_gemm_reduce_impl.hpp" #include "profiler/include/profile_batched_gemm_reduce_impl.hpp"
int main() int main()
{ {
......
test/block_to_ctile_map/test_block_to_ctile_map.cpp
View file @ 7a3b49e5
#include <ck/config.hpp> // SPDX-License-Identifier: MIT
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp" // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include <iostream> #include <iostream>
#include <vector> #include <vector>
#include <gtest/gtest.h>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
using namespace ck; using namespace ck;
......
test/client_app/CMakeLists.txt deleted 100644 → 0
View file @ e07b3d8e
cmake_minimum_required(VERSION 3.15)
project(ck_app)
add_compile_options(-std=c++14)
find_package(composable_kernel 1.0.0 COMPONENTS device_operations host_tensor)
find_package(hip REQUIRED PATHS /opt/rocm)
message(STATUS "Build with HIP ${hip_VERSION}")
add_executable(test_client_app client_app.cpp)
target_link_libraries(test_client_app PRIVATE composable_kernel::device_operations composable_kernel::host_tensor hip::host)
test/client_app/client_app.cpp deleted 100644 → 0
View file @ e07b3d8e
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include <vector>
#include "client_app_impl.hpp"
int main(int argc, char* argv[])
{
if(argc != 25)
{
printf("arg1: tensor operation (conv_fwd: ForwardConvolution)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
printf("arg6: verification (0: no; 1: yes)\n");
printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg9: time kernel (0=n0, 1=yes)\n");
printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
}
const ConvDataType data_type = static_cast<ConvDataType>(std::stoi(argv[2]));
const int in_layout = static_cast<ConvInputLayout>(std::stoi(argv[3]));
const int wei_layout = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
const int out_layout = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const bool time_kernel = std::stoi(argv[9]);
const ck::index_t N = std::stoi(argv[10]);
const ck::index_t K = std::stoi(argv[11]);
const ck::index_t C = std::stoi(argv[12]);
const ck::index_t Y = std::stoi(argv[13]);
const ck::index_t X = std::stoi(argv[14]);
const ck::index_t Hi = std::stoi(argv[15]);
const ck::index_t Wi = std::stoi(argv[16]);
const ck::index_t conv_stride_h = std::stoi(argv[17]);
const ck::index_t conv_stride_w = std::stoi(argv[18]);
const ck::index_t conv_dilation_h = std::stoi(argv[19]);
const ck::index_t conv_dilation_w = std::stoi(argv[20]);
const ck::index_t in_left_pad_h = std::stoi(argv[21]);
const ck::index_t in_left_pad_w = std::stoi(argv[22]);
const ck::index_t in_right_pad_h = std::stoi(argv[23]);
const ck::index_t in_right_pad_w = std::stoi(argv[24]);
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
ck::app::profile_conv_fwd_impl(do_verification,
init_method,
do_log,
time_kernel,
data_type,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
return 1;
}
test/client_app/client_app_impl.hpp deleted 100644 → 0
View file @ e07b3d8e
#pragma once
#include "host_interface.hpp"
enum ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
};
enum ConvInputLayout
{
NCHW, // 0
NHWC, // 1
};
enum ConvWeightLayout
{
KCYX, // 0
KYXC, // 1
};
enum ConvOutputLayout
{
NKHW, // 0
NHWK, // 1
};
void check_hip_error(void)
{
hipError_t err = hipGetLastError();
if(err != hipSuccess)
{
std::cerr << "Error: " << hipGetErrorString(err) << std::endl;
exit(err);
}
}
std::string getDeviceName(int device)
{
struct hipDeviceProp_t prop;
hipGetDeviceProperties(&prop, device);
check_hip_error();
return std::string(prop.name);
}
int getDriver(void)
{
int driver;
hipDriverGetVersion(&driver);
check_hip_error();
return driver;
}
namespace ck {
namespace app {
struct DeviceMem
{
DeviceMem() = delete;
DeviceMem(std::size_t mem_size);
void* GetDeviceBuffer();
void ToDevice(const void* p);
void FromDevice(void* p);
~DeviceMem();
void* mpDeviceBuf;
std::size_t mMemSize;
};
DeviceMem::DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
{
hipGetErrorString(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
}
void* DeviceMem::GetDeviceBuffer() { return mpDeviceBuf; }
void DeviceMem::ToDevice(const void* p)
{
hipGetErrorString(
hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
}
void DeviceMem::FromDevice(void* p)
{
hipGetErrorString(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
}
DeviceMem::~DeviceMem() { hipGetErrorString(hipFree(mpDeviceBuf)); }
void profile_conv_fwd_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
ConvDataType data_type,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
{
const ck::index_t Y = filter_spatial_lengths[0];
const ck::index_t X = filter_spatial_lengths[1];
const ck::index_t Hi = input_spatial_lengths[0];
const ck::index_t Wi = input_spatial_lengths[1];
const ck::index_t Ho = output_spatial_lengths[0];
const ck::index_t Wo = output_spatial_lengths[1];
const auto in_sz = N * C * Hi * Wi;
const auto wei_sz = K * C * Y * X;
const auto out_sz = N * K * Ho * Wo;
using WeiDataType = float;
using InDataType = float;
using OutDataType = float;
app::DeviceMem in_device_buf(sizeof(InDataType) * in_sz);
app::DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_sz);
app::DeviceMem out_device_buf(sizeof(OutDataType) * out_sz);
// data is already on device!
// add device Conv instances
std::vector<DeviceConvFwdPtr_t> conv_ptrs;
if(data_type == F16_F16_F16)
{
add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances_t(conv_ptrs);
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances_t(conv_ptrs);
}
else if(data_type == BF16_BF16_BF16)
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances_t(conv_ptrs);
else if(data_type == F32_F32_F32)
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances_t(conv_ptrs);
else if(data_type == INT8_INT8_INT8)
add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances_t(conv_ptrs);
else
throw std::runtime_error("wrong! Invalid data type");
if(conv_ptrs.empty())
{
throw std::runtime_error("wrong! no device Conv instance found");
}
std::string best_conv_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
int deviceIndex = 0;
hipSetDevice(deviceIndex);
check_hip_error();
StreamConfig stream_config{nullptr, time_kernel};
hipStreamCreate(&stream_config.stream_id_);
check_hip_error();
// profile device Conv instances
for(auto& conv_ptr : conv_ptrs)
{
auto argument_ptr =
conv_ptr.MakeArgumentPointer(static_cast<void*>(in_device_buf.GetDeviceBuffer()),
static_cast<void*>(wei_device_buf.GetDeviceBuffer()),
static_cast<void*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
auto invoker_ptr = conv_ptr.MakeInvokerPointer();
if(conv_ptr.IsSupportedArgument(argument_ptr.get()))
{
std::string conv_name = conv_ptr.GetTypeString();
float ave_time = invoker_ptr->Run(argument_ptr.get(), stream_config);
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo);
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, " << conv_name << std::endl;
if(tflops > best_tflops)
{
best_conv_name = conv_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
}
} // namespace app
} // namespace ck
test/conv2d_bwd_data/conv2d_bwd_data.cpp
View file @ 7a3b49e5
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "config.hpp" #include "config.hpp"
#include "device.hpp" #include "device.hpp"
#include "host_tensor.hpp" #include "host_tensor.hpp"
......
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