Skip to content

GitLab

You need to sign in or sign up before continuing.
Commit 09d4c3a4 authored by illsilin's avatar illsilin
Browse files

merge from public repo

parents 171ed358 8e4c3fb1
Hide whitespace changes
Inline Side-by-side
Showing with 1747 additions and 18 deletions
profiler/src/profile_grouped_conv_fwd.cpp
View file @ 09d4c3a4
......@@ -15,6 +15,7 @@ enum struct ConvLayout
{
GNHWC_GKYXC_GNHWK, // 0
NHWGC_GKYXC_NHWGK, // 1
NGCHW_GKYXC_NGKHW, // 2
};
enum struct ConvDataType
......@@ -54,6 +55,8 @@ static void print_helper_msg()
<< "arg3: indexing data type (0: 32-bit, 1: 64-bit)\n"
<< "arg4: tensor layout (0: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]\n"
<< " 1: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, Ho, Wo, G, K])\n"
<< " 2: Input[N, G, C, Hi, Wi], Weight[G, K, Y, X, C], Output[N, "
"G, K, Ho, Wo]\n"
<< "arg5: verification (0: no, 1: yes)\n"
<< "arg6: initialization (0: no init, 1: integer value, 2: decimal value)\n"
<< "arg7: print tensor value (0: no; 1: yes)\n"
......@@ -111,6 +114,11 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
using GNHWK = ck::tensor_layout::convolution::GNHWK;
using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
//
using NGCHW = ck::tensor_layout::convolution::NGCHW;
using NGKHW = ck::tensor_layout::convolution::NGKHW;
//
using NWGC = ck::tensor_layout::convolution::NWGC;
using NHWGC = ck::tensor_layout::convolution::NHWGC;
......@@ -284,6 +292,17 @@ int profile_grouped_conv_fwd(int argc, char* argv[])
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, INT8{}, INT8{}, INT8{}, INT8{}, INT8{});
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::NGCHW_GKYXC_NGKHW)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, NGCHW{}, GKYXC{}, NGKHW{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, NGCHW{}, GKYXC{}, NGKHW{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
if(data_type == ConvDataType::F32_F32_F32)
......
profiler/src/profile_max_pool2d_bwd.cpp 0 → 100644
View file @ 09d4c3a4
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_max_pool2d_bwd_impl.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
struct maxPoolbwdArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}},
{"wsize", {}},
{"wstride", {}},
{"wdilation", {}},
{"pad1", {}},
{"pad2", {}}};
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_max_pool2d_bwd()
{
std::cout << "arg1: data type (0: fp16; 1: fp32; 3: int8; 5: bf16)\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: input tensor length for NCHW(e.g, --length 2 32 30 30) \n"
<< "--wsize: window size for YX (e.g, --wsize 2 2) \n"
<< "--wstride: window stride for HW (e.g, --wstride 2 2) \n"
<< "--wdilation: window dilation for HW (e.g, --wdilation 1 1) \n"
<< "--pad1: left side of padding in HW (e.g, --pad1 1 1) \n"
<< "--pad2: right side of padding in HW (e.g, --pad2 1 1) \n"
<< "eg: ckProfiler max_pool2d_bwd 0 1 2 0 --length 2 32 30 30 --wsize 2 2 "
"--wstride 2 2 --wdilation 1 1 --pad1 1 1 --pad2 1 1"
<< std::endl;
}
int profile_max_pool2d_bwd(int argc, char* argv[])
{
ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
bool do_verification = true;
int init_method = 2;
bool do_log = false;
bool time_kernel = true;
std::vector<index_t> in_length = {2, 32, 30, 30};
std::vector<index_t> wsize = {2, 2};
std::vector<index_t> wstride = {2, 2};
std::vector<index_t> wdilation = {1, 1};
std::vector<index_t> pad1 = {1, 1};
std::vector<index_t> pad2 = {1, 1};
if(argc != 2 && argc != 33)
{
print_help_max_pool2d_bwd();
return 0;
}
else if(argc == 33)
{
data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
do_verification = std::stoi(argv[3]);
init_method = std::stoi(argv[4]);
do_log = std::stoi(argv[5]);
time_kernel = std::stoi(argv[6]);
// parse the long options
maxPoolbwdArgParser arg_parser;
arg_parser(argc, argv);
in_length = arg_parser.long_opts["length"];
wsize = arg_parser.long_opts["wsize"];
wstride = arg_parser.long_opts["wstride"];
wdilation = arg_parser.long_opts["wdilation"];
pad1 = arg_parser.long_opts["pad1"];
pad2 = arg_parser.long_opts["pad2"];
}
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using I8 = int8_t;
using I32 = int32_t;
if(data_type == ck::DataTypeEnum::Half)
{
ck::profiler::profile_max_pool2d_bwd_impl<F16, F16, I32, F16, F16, false>(do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
wdilation,
pad1,
pad2);
}
else if(data_type == ck::DataTypeEnum::BFloat16)
{
ck::profiler::profile_max_pool2d_bwd_impl<BF16, BF16, I32, BF16, BF16, false>(
do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
wdilation,
pad1,
pad2);
}
else if(data_type == ck::DataTypeEnum::Float)
{
ck::profiler::profile_max_pool2d_bwd_impl<F32, F32, I32, F32, F32, false>(do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
wdilation,
pad1,
pad2);
}
else if(data_type == ck::DataTypeEnum::Int8)
{
ck::profiler::profile_max_pool2d_bwd_impl<I8, I8, I32, I8, I8, false>(do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
wdilation,
pad1,
pad2);
}
else
{
throw std::runtime_error("not implemented yet");
}
return 0;
}
REGISTER_PROFILER_OPERATION("max_pool2d_bwd", "max_pool2d bwd", profile_max_pool2d_bwd);
profiler/src/profile_max_pool3d_fwd.cpp → profiler/src/profile_max_pool2d_fwd.cpp
View file @ 09d4c3a4
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_pool3d_fwd_impl.hpp"
#include "profiler/profile_pool2d_fwd_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
......@@ -49,49 +49,58 @@ struct maxPoolFwdArgParser
}
};
void print_help_max_pool3d_fwd()
enum struct PoolDataType
{
std::cout << "arg1: data type (0: fp16; 1: fp32; 5: bf16)\n"
F32 = 0,
BF16,
F16,
INT8,
F8,
};
void print_help_max_pool2d_fwd()
{
std::cout << "arg1: data type (0: fp16; 1: fp32; 2: bf16; 3: int8; 4: fp8)\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"
<< "arg6: return index (0=no, 1=yes)\n"
<< "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n"
<< "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n"
<< "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n"
<< "--wdilation: window dilation for DHW (e.g, --wdilation 1 1 1) \n"
<< "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n"
<< "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n"
<< "eg: ckProfiler max_pool3d_fwd 0 1 2 0 1 0 --length 2 32 30 30 30 --wsize 2 2 2 "
"--wstride 2 2 2 --wdilation 1 1 1 --pad1 1 1 1 --pad2 1 1 1"
<< "--length: input tensor length for NCHW(e.g, --length 2 32 30 30) \n"
<< "--wsize: window size for YX (e.g, --wsize 2 2) \n"
<< "--wstride: window stride for HW (e.g, --wstride 2 2) \n"
<< "--wdilation: window dilation for HW (e.g, --wdilation 1 1) \n"
<< "--pad1: left side of padding in HW (e.g, --pad1 1 1) \n"
<< "--pad2: right side of padding in HW (e.g, --pad2 1 1) \n"
<< "eg: ckProfiler max_pool2d_fwd 0 1 2 0 1 0 --length 2 32 30 30 --wsize 2 2"
"--wstride 2 2 --wdilation 1 1 --pad1 1 1 --pad2 1 1"
<< std::endl;
}
int profile_max_pool3d_fwd(int argc, char* argv[])
int profile_max_pool2d_fwd(int argc, char* argv[])
{
ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
bool do_verification = true;
int init_method = 0;
bool do_log = false;
bool time_kernel = true;
bool return_index = false;
std::vector<index_t> in_length = {2, 32, 30, 30, 30};
std::vector<index_t> wsize = {2, 2, 2};
std::vector<index_t> wstride = {2, 2, 2};
std::vector<index_t> wdilation = {1, 1, 1};
std::vector<index_t> pad1 = {1, 1, 1};
std::vector<index_t> pad2 = {1, 1, 1};
if(argc != 2 && argc != 34)
PoolDataType data_type = PoolDataType::F32;
bool do_verification = true;
int init_method = 0;
bool do_log = false;
bool time_kernel = true;
bool return_index = false;
std::vector<index_t> in_length = {2, 32, 30, 30};
std::vector<index_t> wsize = {2, 2};
std::vector<index_t> wstride = {2, 2};
std::vector<index_t> wdilation = {1, 1};
std::vector<index_t> pad1 = {1, 1};
std::vector<index_t> pad2 = {1, 1};
if(argc != 2 && argc != 28)
{
print_help_max_pool3d_fwd();
print_help_max_pool2d_fwd();
return 0;
}
else if(argc == 34)
else if(argc == 28)
{
data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
data_type = static_cast<PoolDataType>(std::stoi(argv[2]));
do_verification = std::stoi(argv[3]);
init_method = std::stoi(argv[4]);
do_log = std::stoi(argv[5]);
......@@ -109,32 +118,22 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
pad2 = arg_parser.long_opts["pad2"];
}
#ifdef CK_ENABLE_FP16
using F16 = ck::half_t;
#endif
#ifdef CK_ENABLE_BF16
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
#endif
#ifdef CK_ENABLE_FP32
using F32 = float;
#endif
using I32 = int32_t;
using NDHWC = ck::tensor_layout::convolution::NDHWC;
#if 1
using F32 = float;
using I32 = int32_t;
using F8 = ck::f8_t;
using I8 = int8_t;
using NHWC = ck::tensor_layout::convolution::NHWC;
constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
#else
constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
#endif
if(false)
;
#ifdef CK_ENABLE_FP16
else if(data_type == ck::DataTypeEnum::Half)
if(data_type == PoolDataType::F16)
{
if(return_index)
{
ck::profiler::
profile_pool3d_fwd_impl<F16, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false, true>(
profile_pool2d_fwd_impl<F16, F16, F16, I32, NHWC, NHWC, ReduceOpId, false, true>(
do_verification,
init_method,
do_log,
......@@ -145,9 +144,11 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
wdilation,
pad1,
pad2);
}
else
{
ck::profiler::
profile_pool3d_fwd_impl<F16, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false, false>(
profile_pool2d_fwd_impl<F16, F16, F16, I32, NHWC, NHWC, ReduceOpId, false, false>(
do_verification,
init_method,
do_log,
......@@ -158,37 +159,33 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
wdilation,
pad1,
pad2);
}
}
#endif
#ifdef CK_ENABLE_BF16
else if(data_type == ck::DataTypeEnum::BFloat16)
else if(data_type == PoolDataType::BF16)
{
if(return_index)
ck::profiler::profile_pool3d_fwd_impl<BF16,
BF16,
BF16,
I32,
NDHWC,
NDHWC,
ReduceOpId,
false,
true>(do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
wdilation,
pad1,
pad2);
{
ck::profiler::
profile_pool2d_fwd_impl<BF16, BF16, BF16, I32, NHWC, NHWC, ReduceOpId, false, true>(
do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
wdilation,
pad1,
pad2);
}
else
ck::profiler::profile_pool3d_fwd_impl<BF16,
{
ck::profiler::profile_pool2d_fwd_impl<BF16,
BF16,
BF16,
I32,
NDHWC,
NDHWC,
NHWC,
NHWC,
ReduceOpId,
false,
false>(do_verification,
......@@ -201,14 +198,14 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
wdilation,
pad1,
pad2);
}
}
#endif
#ifdef CK_ENABLE_FP32
else if(data_type == ck::DataTypeEnum::Float)
else if(data_type == PoolDataType::F32)
{
if(return_index)
{
ck::profiler::
profile_pool3d_fwd_impl<F32, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false, true>(
profile_pool2d_fwd_impl<F32, F32, F32, I32, NHWC, NHWC, ReduceOpId, false, true>(
do_verification,
init_method,
do_log,
......@@ -219,9 +216,11 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
wdilation,
pad1,
pad2);
}
else
{
ck::profiler::
profile_pool3d_fwd_impl<F32, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false, false>(
profile_pool2d_fwd_impl<F32, F32, F32, I32, NHWC, NHWC, ReduceOpId, false, false>(
do_verification,
init_method,
do_log,
......@@ -232,8 +231,74 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
wdilation,
pad1,
pad2);
}
}
else if(data_type == PoolDataType::INT8)
{
if(return_index)
{
ck::profiler::
profile_pool2d_fwd_impl<I8, I8, F32, I32, NHWC, NHWC, ReduceOpId, false, true>(
do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
wdilation,
pad1,
pad2);
}
else
{
ck::profiler::
profile_pool2d_fwd_impl<I8, I8, F32, I32, NHWC, NHWC, ReduceOpId, false, false>(
do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
wdilation,
pad1,
pad2);
}
}
else if(data_type == PoolDataType::F8)
{
if(return_index)
{
ck::profiler::
profile_pool2d_fwd_impl<F8, F8, F32, I32, NHWC, NHWC, ReduceOpId, false, true>(
do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
wdilation,
pad1,
pad2);
}
else
{
ck::profiler::
profile_pool2d_fwd_impl<F8, F8, F32, I32, NHWC, NHWC, ReduceOpId, false, false>(
do_verification,
init_method,
do_log,
time_kernel,
in_length,
wsize,
wstride,
wdilation,
pad1,
pad2);
}
}
#endif
else
{
throw std::runtime_error("not implemented yet");
......@@ -242,4 +307,4 @@ int profile_max_pool3d_fwd(int argc, char* argv[])
return 0;
}
REGISTER_PROFILER_OPERATION("max_pool3d_fwd", "max_pool3d fwd", profile_max_pool3d_fwd);
REGISTER_PROFILER_OPERATION("max_pool2d_fwd", "max_pool2d fwd", profile_max_pool2d_fwd);
profiler/src/profile_pool3d_fwd.cpp 0 → 100644
View file @ 09d4c3a4
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/data_type_enum.hpp"
#include "profiler/profile_pool3d_fwd_impl.hpp"
#include "profiler_operation_registry.hpp"
using ck::index_t;
struct poolFwdArgParser
{
std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}},
{"wsize", {}},
{"wstride", {}},
{"wdilation", {}},
{"pad1", {}},
{"pad2", {}}};
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_pool3d_fwd()
{
std::cout << "arg1: data type (0: fp16; 1: fp32; 3: int8; 5: bf16; 7: fp8)\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"
<< "arg6: return index (0=no, 1=yes)\n"
<< "arg7: reduce op (0: max; 1: avg)\n"
<< "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n"
<< "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n"
<< "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n"
<< "--wdilation: window dilation for DHW (e.g, --wdilation 1 1 1) \n"
<< "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n"
<< "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n"
<< "eg: ckProfiler pool3d_fwd 0 1 2 0 1 0 --length 2 32 30 30 30 --wsize 2 2 2 "
"--wstride 2 2 2 --wdilation 1 1 1 --pad1 1 1 1 --pad2 1 1 1"
<< std::endl;
}
int profile_pool3d_fwd(int argc, char* argv[])
{
ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
ck::profiler::PoolFwdInputParams in_params{true, 0, false, true, false, 0};
ck::profiler::PoolFwdKernelParams kernel_params{
{2, 32, 30, 30, 30}, {2, 2, 2}, {2, 2, 2}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}};
if(argc != 2 && argc != 35)
{
print_help_pool3d_fwd();
return 0;
}
else if(argc == 35)
{
data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
in_params.do_verification = std::stoi(argv[3]);
in_params.init_method = std::stoi(argv[4]);
in_params.do_log = std::stoi(argv[5]);
in_params.time_kernel = std::stoi(argv[6]);
in_params.return_index = std::stoi(argv[7]);
in_params.reduce_op = std::stoi(argv[8]);
// parse the long options
poolFwdArgParser arg_parser;
arg_parser(argc, argv);
kernel_params.in_length = arg_parser.long_opts["length"];
kernel_params.window_spatial_lengths = arg_parser.long_opts["wsize"];
kernel_params.window_strides = arg_parser.long_opts["wstride"];
kernel_params.window_dilations = arg_parser.long_opts["wdilation"];
kernel_params.input_left_pads = arg_parser.long_opts["pad1"];
kernel_params.input_right_pads = arg_parser.long_opts["pad2"];
}
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using I8 = int8_t;
using I32 = int32_t;
using F8 = ck::f8_t;
using NDHWC = ck::tensor_layout::convolution::NDHWC;
if(data_type == ck::DataTypeEnum::Half)
{
if(in_params.reduce_op == 1)
{
ck::profiler::profile_pool3d_fwd_impl<F16,
F16,
F32,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::AVG,
false,
false>(in_params, kernel_params);
}
else
{ // reduce_op == 0
if(in_params.return_index)
{
ck::profiler::profile_pool3d_fwd_impl<F16,
F16,
F16,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::MAX,
false,
true>(in_params, kernel_params);
}
else
{
ck::profiler::profile_pool3d_fwd_impl<F16,
F16,
F16,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::MAX,
false,
false>(in_params, kernel_params);
}
}
}
else if(data_type == ck::DataTypeEnum::BFloat16)
{
if(in_params.reduce_op == 1)
{
ck::profiler::profile_pool3d_fwd_impl<BF16,
BF16,
F32,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::AVG,
false,
false>(in_params, kernel_params);
}
else
{ // reduce_op == 0
if(in_params.return_index)
{
ck::profiler::profile_pool3d_fwd_impl<BF16,
BF16,
BF16,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::MAX,
false,
true>(in_params, kernel_params);
}
else
{
ck::profiler::profile_pool3d_fwd_impl<BF16,
BF16,
BF16,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::MAX,
false,
false>(in_params, kernel_params);
}
}
}
else if(data_type == ck::DataTypeEnum::Float)
{
if(in_params.reduce_op == 1)
{
ck::profiler::profile_pool3d_fwd_impl<F32,
F32,
F32,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::AVG,
false,
false>(in_params, kernel_params);
}
else
{ // reduce_op == 0
if(in_params.return_index)
{
ck::profiler::profile_pool3d_fwd_impl<F32,
F32,
F32,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::MAX,
false,
true>(in_params, kernel_params);
}
else
{
ck::profiler::profile_pool3d_fwd_impl<F32,
F32,
F32,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::MAX,
false,
false>(in_params, kernel_params);
}
}
}
else if(data_type == ck::DataTypeEnum::Float8)
{
if(in_params.reduce_op == 1)
{
return ck::profiler::profile_pool3d_fwd_impl<F8,
F8,
F32,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::AVG,
false,
false>(in_params, kernel_params);
}
else
{ // reduce_op == 0
if(in_params.return_index)
{
return ck::profiler::profile_pool3d_fwd_impl<F8,
F8,
F8,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::MAX,
false,
true>(in_params, kernel_params);
}
else
{
return ck::profiler::profile_pool3d_fwd_impl<F8,
F8,
F8,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::MAX,
false,
false>(in_params, kernel_params);
}
}
}
else if(data_type == ck::DataTypeEnum::Int8)
{
if(in_params.reduce_op == 1)
{
return ck::profiler::profile_pool3d_fwd_impl<I8,
I8,
I32,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::AVG,
false,
false>(in_params, kernel_params);
}
else
{ // reduce_op == 0
if(in_params.return_index)
{
return ck::profiler::profile_pool3d_fwd_impl<I8,
I8,
I8,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::MAX,
false,
true>(in_params, kernel_params);
}
else
{
return ck::profiler::profile_pool3d_fwd_impl<I8,
I8,
I8,
I32,
NDHWC,
NDHWC,
ck::ReduceTensorOp::MAX,
false,
false>(in_params, kernel_params);
}
}
}
else
{
throw std::runtime_error("not implemented yet");
}
return 0;
}
REGISTER_PROFILER_OPERATION("pool3d_fwd", "pool3d fwd", profile_pool3d_fwd);
script/cmake-ck-dev.sh
View file @ 09d4c3a4
......@@ -7,8 +7,10 @@ MY_PROJECT_SOURCE=$1
if [ $# -ge 2 ] ; then
GPU_TARGETS=$2
REST_ARGS=${@:3}
else
GPU_TARGETS="gfx908;gfx90a;gfx940"
REST_ARGS=
fi
cmake \
......@@ -20,4 +22,5 @@ cmake
-D GPU_TARGETS=$GPU_TARGETS \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D USE_BITINT_EXTENSION_INT4=OFF \
$REST_ARGS \
${MY_PROJECT_SOURCE}
script/cmake-ck-release.sh
View file @ 09d4c3a4
......@@ -7,8 +7,10 @@ MY_PROJECT_SOURCE=$1
if [ $# -ge 2 ] ; then
GPU_TARGETS=$2
REST_ARGS=${@:3}
else
GPU_TARGETS="gfx908;gfx90a;gfx940"
REST_ARGS=
fi
cmake \
......@@ -20,5 +22,6 @@ cmake
-D GPU_TARGETS=$GPU_TARGETS \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
-D USE_BITINT_EXTENSION_INT4=OFF \
$REST_ARGS \
${MY_PROJECT_SOURCE}
script/convert_miopen_driver_to_profiler.py
View file @ 09d4c3a4
......@@ -28,6 +28,8 @@ def parse_layouts(args):
args.in_layout == "NCDHW":
if args.ck_profier_op == "grouped_conv_bwd_weight":
args.layout = 3
elif args.ck_profier_op == "grouped_conv_fwd":
args.layout = 2
else:
print('Not supported layout for this op')
exit(1)
......
test/CMakeLists.txt
View file @ 09d4c3a4
......@@ -173,6 +173,7 @@ function(add_gtest_executable TEST_NAME)
endfunction()
add_compile_options(-Wno-c++20-extensions)
add_subdirectory(ck_tile)
add_subdirectory(magic_number_division)
add_subdirectory(space_filling_curve)
add_subdirectory(conv_util)
......
test/ck_tile/CMakeLists.txt 0 → 100644
View file @ 09d4c3a4
add_subdirectory(image_to_column)
test/ck_tile/image_to_column/CMakeLists.txt 0 → 100644
View file @ 09d4c3a4
# Currently ck_tile is only built on gfx9
if(GPU_TARGETS MATCHES "gfx9")
add_gtest_executable(test_tile_image_to_column test_tile_image_to_column.cpp)
endif()
test/ck_tile/image_to_column/test_tile_image_to_column.cpp 0 → 100644
View file @ 09d4c3a4
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <algorithm>
#include <gtest/gtest.h>
#include "ck_tile/host.hpp"
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/image_to_column.hpp"
// Host API implementation
template <typename DataType>
class TestCkTileImageToColumn : public ::testing::Test
{
static constexpr ck_tile::index_t VectorSize = 1;
static constexpr ck_tile::index_t NDimSpatial = 2;
protected:
void Run(const ck_tile::conv::ConvParam conv_params)
{
using ImLayout = ck_tile::tensor_layout::convolution::NHWGC;
const auto G = conv_params.G_;
const auto N = conv_params.N_;
const auto C = conv_params.C_;
const ck_tile::long_index_t NDoHoWo =
N * std::accumulate(conv_params.output_spatial_lengths_.begin(),
std::next(conv_params.output_spatial_lengths_.begin(), NDimSpatial),
1,
std::multiplies<>());
const ck_tile::long_index_t CZYX =
C * std::accumulate(conv_params.filter_spatial_lengths_.begin(),
std::next(conv_params.filter_spatial_lengths_.begin(), NDimSpatial),
1,
std::multiplies<>());
const auto in_desc =
ck_tile::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ImLayout>(
conv_params);
const auto out_desc = ck_tile::HostTensorDescriptor({G, NDoHoWo, CZYX});
// host verify
ck_tile::HostTensor<DataType> in(in_desc);
ck_tile::HostTensor<DataType> out_device(out_desc);
ck_tile::HostTensor<DataType> out_host(out_desc);
std::cout << "input: " << in.mDesc << std::endl;
std::cout << "output: " << out_device.mDesc << std::endl;
ck_tile::FillUniformDistributionIntegerValue<DataType>{-5.f, 5.f}(in);
ck_tile::DeviceMem in_device_buf(in.get_element_space_size_in_bytes());
ck_tile::DeviceMem out_device_buf(out_device.get_element_space_size_in_bytes());
in_device_buf.ToDevice(in.data());
using thread_tile = ck_tile::sequence<4, 4>;
using warp_tile = ck_tile::sequence<8, 128>;
using block_tile = ck_tile::sequence<32, 128>;
using Shape = ck_tile::TileImageToColumnShape<thread_tile, warp_tile, block_tile>;
using PipelineProblem = ck_tile::BlockImageToColumnProblem<DataType,
DataType,
Shape,
NDimSpatial,
VectorSize,
VectorSize>;
using Kernel = ck_tile::ImageToColumn<PipelineProblem>;
auto kargs = Kernel::MakeKargs(
in_device_buf.GetDeviceBuffer(),
out_device_buf.GetDeviceBuffer(),
G,
N,
C,
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(
conv_params.input_spatial_lengths_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(
conv_params.filter_spatial_lengths_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(
conv_params.output_spatial_lengths_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial + 3>(in_desc.get_strides()),
ck_tile::to_array<ck_tile::long_index_t, 3>(out_desc.get_strides()),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.conv_filter_strides_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(
conv_params.conv_filter_dilations_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.input_left_pads_),
ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.input_right_pads_));
const dim3 grids = Kernel::GridSize(
kargs.N * kargs.output_spatial_lengths[0] * kargs.output_spatial_lengths[1],
kargs.filter_spatial_lengths[0] * kargs.filter_spatial_lengths[1] * kargs.C,
kargs.G);
constexpr dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 2;
ck_tile::launch_kernel(
ck_tile::stream_config{},
ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
// reference
ck_tile::reference_im2col<DataType, DataType, NDimSpatial>(in, out_host, conv_params);
out_device_buf.FromDevice(out_device.data());
bool pass = ck_tile::check_err(out_device, out_host);
EXPECT_TRUE(pass);
}
};
class TestCkTileImageToColumnFloat : public TestCkTileImageToColumn<float>
{
};
class TestCkTileImageToColumnHalf : public TestCkTileImageToColumn<ck_tile::half_t>
{
};
TEST_F(TestCkTileImageToColumnFloat, TestCorrectness)
{
this->Run({2, 2, 4, 1, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->Run({2, 2, 64, 1, 64, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->Run({2, 1, 64, 1, 64, {1, 1}, {7, 7}, {3, 3}, {1, 1}, {0, 0}, {0, 0}});
this->Run({2, 1, 64, 1, 64, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->Run({2, 2, 64, 1, 64, {3, 3}, {28, 28}, {2, 2}, {2, 2}, {1, 1}, {1, 1}});
}
TEST_F(TestCkTileImageToColumnHalf, TestCorrectness)
{
this->Run({2, 2, 4, 1, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->Run({2, 2, 64, 1, 64, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
this->Run({2, 1, 64, 1, 64, {1, 1}, {7, 7}, {3, 3}, {1, 1}, {0, 0}, {0, 0}});
this->Run({2, 1, 64, 1, 64, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
this->Run({2, 2, 64, 1, 64, {3, 3}, {28, 28}, {2, 2}, {2, 2}, {1, 1}, {1, 1}});
}
test/gemm_universal/test_gemm_universal_ut_cases.inc
View file @ 09d4c3a4
......@@ -28,6 +28,38 @@ TYPED_TEST(TestGemmUniversal_MK_NK, SmallM)
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
TYPED_TEST(TestGemmUniversal_KM_KN, SmallM)
{
std::vector<int> Ms{1, 2, 3, 4, 5, 6};
constexpr int N = 512;
constexpr int K = 320;
constexpr int StrideB = N;
constexpr int StrideC = N;
for(int M : Ms)
{
int StrideA = M;
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
}
TYPED_TEST(TestGemmUniversal_KM_NK, SmallM)
{
std::vector<int> Ms{1, 2, 3, 4, 5, 6};
constexpr int N = 512;
constexpr int K = 320;
constexpr int StrideB = N;
constexpr int StrideC = N;
for(int M : Ms)
{
int StrideA = M;
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
}
TYPED_TEST(TestGemmUniversal_MK_KN, MidLargeM)
{
std::vector<int> Ms{127, 255, 312, 799, 1573};
......@@ -56,6 +88,38 @@ TYPED_TEST(TestGemmUniversal_MK_NK, MidLargeM)
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
TYPED_TEST(TestGemmUniversal_KM_KN, MidLargeM)
{
std::vector<int> Ms{127, 255, 312, 799, 1573};
constexpr int N = 512;
constexpr int K = 320;
constexpr int StrideB = N;
constexpr int StrideC = N;
for(int M : Ms)
{
int StrideA = M;
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
}
TYPED_TEST(TestGemmUniversal_KM_NK, MidLargeM)
{
std::vector<int> Ms{127, 255, 312, 799, 1573};
constexpr int N = 512;
constexpr int K = 320;
constexpr int StrideB = N;
constexpr int StrideC = N;
for(int M : Ms)
{
int StrideA = M;
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
}
TYPED_TEST(TestGemmUniversal_MK_KN, PaddK)
{
std::vector<int> Ms{127};
......@@ -84,6 +148,38 @@ TYPED_TEST(TestGemmUniversal_MK_NK, PaddK)
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
TYPED_TEST(TestGemmUniversal_KM_KN, PaddK)
{
std::vector<int> Ms{127};
constexpr int N = 512;
constexpr int K = 437;
constexpr int StrideB = N;
constexpr int StrideC = N;
for(int M : Ms)
{
int StrideA = M;
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
}
TYPED_TEST(TestGemmUniversal_KM_NK, PaddK)
{
std::vector<int> Ms{127};
constexpr int N = 512;
constexpr int K = 437;
constexpr int StrideB = N;
constexpr int StrideC = N;
for(int M : Ms)
{
int StrideA = M;
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
}
TYPED_TEST(TestGemmUniversal_MK_KN, Regular)
{
std::vector<int> Ms{512};
......@@ -111,3 +207,35 @@ TYPED_TEST(TestGemmUniversal_MK_NK, Regular)
for(int M : Ms)
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
TYPED_TEST(TestGemmUniversal_KM_KN, Regular)
{
std::vector<int> Ms{512};
constexpr int N = 512;
constexpr int K = 512;
constexpr int StrideB = N;
constexpr int StrideC = N;
for(int M : Ms)
{
int StrideA = M;
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
}
TYPED_TEST(TestGemmUniversal_KM_NK, Regular)
{
std::vector<int> Ms{512};
constexpr int N = 512;
constexpr int K = 512;
constexpr int StrideB = N;
constexpr int StrideC = N;
for(int M : Ms)
{
int StrideA = M;
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
}
test/gemm_universal/test_gemm_universal_xdl.cpp
View file @ 09d4c3a4
......@@ -40,6 +40,18 @@ class TestGemmUniversal_MK_NK
{
};
template <typename Tuple>
class TestGemmUniversal_KM_KN
: public ck::test::TestGemmUniversal<typename tuple_concat<std::tuple<Col, Row>, Tuple>::type>
{
};
template <typename Tuple>
class TestGemmUniversal_KM_NK
: public ck::test::TestGemmUniversal<typename tuple_concat<std::tuple<Col, Col>, Tuple>::type>
{
};
// clang-format off
using KernelTypes_MK_KN = ::testing::Types<
// ADataType, BDataType, ComputeDataType, CDataType
......@@ -61,9 +73,22 @@ using KernelTypes_MK_NK = ::testing::Types<
#endif
std::tuple< BF16, BF16, BF16, BF16>
>;
using KernelTypes_KM_NK = ::testing::Types<
// ADataType, BDataType, ComputeDataType, CDataType
std::tuple< BF16, BF16, BF16, BF16>
>;
using KernelTypes_KM_KN = ::testing::Types<
// ADataType, BDataType, ComputeDataType, CDataType
std::tuple< BF16, BF16, BF16, BF16>
>;
// clang-format on
TYPED_TEST_SUITE(TestGemmUniversal_MK_KN, KernelTypes_MK_KN);
TYPED_TEST_SUITE(TestGemmUniversal_MK_NK, KernelTypes_MK_NK);
TYPED_TEST_SUITE(TestGemmUniversal_KM_KN, KernelTypes_KM_KN);
TYPED_TEST_SUITE(TestGemmUniversal_KM_NK, KernelTypes_KM_NK);
#include "test_gemm_universal_ut_cases.inc"
test/grouped_convnd_fwd/test_grouped_convnd_fwd.cpp
View file @ 09d4c3a4
......@@ -62,7 +62,9 @@ using KernelTypes2d = ::testing::Types<std::tuple<float, GNHWC, GKYXC, GNHWK>,
std::tuple<float, NHWGC, GKYXC, NHWGK>,
std::tuple<ck::half_t, NHWGC, GKYXC, NHWGK>,
std::tuple<ck::bhalf_t, NHWGC, GKYXC, NHWGK>,
std::tuple<int8_t, NHWGC, GKYXC, NHWGK>>;
std::tuple<int8_t, NHWGC, GKYXC, NHWGK>,
std::tuple<float, NGCHW, GKYXC, NGKHW>,
std::tuple<ck::half_t, NGCHW, GKYXC, NGKHW>>;
using KernelTypes3d = ::testing::Types<std::tuple<float, GNDHWC, GKZYXC, GNDHWK>,
std::tuple<ck::half_t, GNDHWC, GKZYXC, GNDHWK>,
......
test/pool/CMakeLists.txt
View file @ 09d4c3a4
......@@ -4,13 +4,25 @@ add_gtest_executable(test_avg_pool3d_bwd test_avg_pool3d_bwd.cpp)
add_gtest_executable(test_max_pool3d_bwd test_max_pool3d_bwd.cpp)
add_gtest_executable(test_avg_pool3d_fwd test_avg_pool3d_fwd.cpp)
add_gtest_executable(test_max_pool3d_fwd test_max_pool3d_fwd.cpp)
add_gtest_executable(test_avg_pool2d_bwd test_avg_pool2d_bwd.cpp)
add_gtest_executable(test_max_pool2d_bwd test_max_pool2d_bwd.cpp)
add_gtest_executable(test_avg_pool2d_fwd test_avg_pool2d_fwd.cpp)
add_gtest_executable(test_max_pool2d_fwd test_max_pool2d_fwd.cpp)
target_link_libraries(test_avg_pool3d_bwd PRIVATE utility device_avg_pool3d_bwd_instance)
target_link_libraries(test_avg_pool2d_bwd PRIVATE utility device_avg_pool2d_bwd_instance)
target_link_libraries(test_max_pool2d_bwd PRIVATE utility device_max_pool_bwd_instance)
target_link_libraries(test_max_pool3d_bwd PRIVATE utility device_max_pool_bwd_instance)
target_link_libraries(test_avg_pool3d_fwd PRIVATE utility device_pool3d_fwd_instance)
target_link_libraries(test_max_pool3d_fwd PRIVATE utility device_pool3d_fwd_instance)
target_link_libraries(test_avg_pool2d_fwd PRIVATE utility device_pool2d_fwd_instance)
target_link_libraries(test_max_pool2d_fwd PRIVATE utility device_pool2d_fwd_instance)
add_dependencies(test_pool test_avg_pool3d_bwd)
add_dependencies(test_pool test_max_pool3d_bwd)
add_dependencies(test_pool test_avg_pool3d_fwd)
add_dependencies(test_pool test_max_pool3d_fwd)
add_dependencies(test_pool test_avg_pool2d_bwd)
add_dependencies(test_pool test_max_pool2d_bwd)
add_dependencies(test_pool test_avg_pool2d_fwd)
add_dependencies(test_pool test_max_pool2d_fwd)
test/pool/test_avg_pool2d_bwd.cpp 0 → 100644
View file @ 09d4c3a4
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "profiler/profile_avg_pool2d_bwd_impl.hpp"
#include "test_pool_fwd_common.hpp"
template <typename T>
class AvgPool2dBWDTest : public ::testing::Test
{
protected:
using InDataType = std::tuple_element_t<0, T>;
using OutDataType = std::tuple_element_t<1, T>;
static std::vector<PoolingParam> params;
void Run()
{
for(auto param : this->params)
{
bool success =
ck::profiler::profile_avg_pool2d_bwd_impl<InDataType, OutDataType, NHWC, NHWC>(
true,
2,
false,
false,
param.length_,
param.window_spatial_lengths_,
param.window_strides_,
param.window_dilations_,
param.input_left_pads_,
param.input_right_pads_);
EXPECT_TRUE(success);
}
}
};
template <typename T>
std::vector<PoolingParam> AvgPool2dBWDTest<T>::params = {
{{1, 1, 1, 1}, {1, 1}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
{{1, 1, 64, 64}, {64, 64}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
{{1, 5, 7, 7}, {2, 2}, {2, 2}, {1, 1}, {2, 2}, {0, 0}},
{{1, 1, 8, 8}, {2, 2}, {2, 2}, {1, 1}, {2, 2}, {0, 0}},
{{1, 1, 8, 8}, {2, 2}, {1, 1}, {1, 1}, {1, 1}, {0, 0}},
{{2, 32, 30, 30}, {2, 2}, {2, 2}, {1, 1}, {1, 1}, {1, 1}},
{{1, 2, 30, 30}, {2, 2}, {2, 2}, {1, 1}, {0, 0}, {0, 0}}};
using Avg_Pool_2D_f32_types = ::testing::Types<std::tuple<F32, F32>>;
using Avg_Pool_2D_int8_types = ::testing::Types<std::tuple<I8, I8>>;
using Avg_Pool_2D_f16_types = ::testing::Types<std::tuple<F16, F16>>;
using Avg_Pool_2D_bf16_types = ::testing::Types<std::tuple<BF16, BF16>>;
using Avg_Pool_2D_f8_types = ::testing::Types<std::tuple<F8, F8>>;
template <typename TType>
class AvgPool2D_f32 : public AvgPool2dBWDTest<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP32)
{
GTEST_SKIP() << "Skipping AvgPool2D_f32 tests because CK_ENABLE_FP32 is not enabled";
}
}
};
template <typename TType>
class AvgPool2D_int8 : public AvgPool2dBWDTest<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_INT8)
{
GTEST_SKIP() << "Skipping AvgPool2D_int8 tests because CK_ENABLE_INT8 is not enabled";
}
}
};
template <typename TType>
class AvgPool2D_f16 : public AvgPool2dBWDTest<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP16)
{
GTEST_SKIP() << "Skipping AvgPool2D_f16 because CK_ENABLE_FP16 is not enabled";
}
}
};
template <typename TType>
class AvgPool2D_bf16 : public AvgPool2dBWDTest<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_BF16)
{
GTEST_SKIP() << "Skipping AvgPool2D_bf16 tests because CK_ENABLE_BF16 is not enabled";
}
}
};
template <typename TType>
class AvgPool2D_f8 : public AvgPool2dBWDTest<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP8)
{
GTEST_SKIP() << "Skipping AvgPool2D_f8 tests because CK_ENABLE_FP8 is not enabled";
}
}
};
TYPED_TEST_SUITE(AvgPool2D_f32, Avg_Pool_2D_f32_types);
TYPED_TEST_SUITE(AvgPool2D_int8, Avg_Pool_2D_int8_types);
TYPED_TEST_SUITE(AvgPool2D_f16, Avg_Pool_2D_f16_types);
TYPED_TEST_SUITE(AvgPool2D_bf16, Avg_Pool_2D_bf16_types);
TYPED_TEST_SUITE(AvgPool2D_f8, Avg_Pool_2D_f8_types);
TYPED_TEST(AvgPool2D_f32, AvgPool2DTest_f32) { this->Run(); }
TYPED_TEST(AvgPool2D_int8, AvgPool2DTest_int8) { this->Run(); }
TYPED_TEST(AvgPool2D_f16, AvgPool2DTest_f16) { this->Run(); }
TYPED_TEST(AvgPool2D_bf16, AvgPool2DTest_bf16) { this->Run(); }
TYPED_TEST(AvgPool2D_f8, AvgPool2DTest_f8) { this->Run(); }
test/pool/test_avg_pool2d_fwd.cpp 0 → 100644
View file @ 09d4c3a4
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "profiler/profile_pool2d_fwd_impl.hpp"
#include "test_pool_fwd_common.hpp"
template <typename Tuple>
class TestAvgPool2dFwd : public ::testing::Test
{
protected:
using InDataType = std::tuple_element_t<0, Tuple>;
using OutDataType = std::tuple_element_t<1, Tuple>;
using ComputeDataType = std::tuple_element_t<2, Tuple>;
using IndexDataType = std::tuple_element_t<3, Tuple>;
static std::vector<PoolingParam> params;
void Run()
{
for(auto param : params)
{
bool success =
ck::profiler::profile_pool2d_fwd_impl<InDataType,
OutDataType,
ComputeDataType,
IndexDataType,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::NHWC,
ck::ReduceTensorOp::AVG,
false,
false>(true,
2,
false,
false,
param.length_,
param.window_spatial_lengths_,
param.window_strides_,
param.window_dilations_,
param.input_left_pads_,
param.input_right_pads_);
EXPECT_TRUE(success);
}
}
};
template <typename T>
std::vector<PoolingParam> TestAvgPool2dFwd<T>::params = {
{{{1, 1, 1, 1}, {1, 1}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
{{2, 16, 64, 64}, {64, 64}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
{{2, 16, 64, 64}, {4, 4}, {4, 4}, {2, 2}, {0, 0}, {0, 0}},
{{2, 32, 30, 30}, {2, 2}, {2, 2}, {1, 1}, {1, 1}, {1, 1}}}};
using AvgPool2D_F32_Types =
::testing::Types<std::tuple<F32, F32, F32, I32>, std::tuple<F32, F32, F32, I32>>;
using AvgPool2D_F16_Types =
::testing::Types<std::tuple<F16, F16, F32, I32>, std::tuple<F16, F16, F32, I32>>;
using AvgPool2D_BF16_Types =
::testing::Types<std::tuple<I8, I8, F32, I32>, std::tuple<BF16, BF16, F32, I32>>;
using AvgPool2D_I8_Types =
::testing::Types<std::tuple<I8, I8, F32, I32>, std::tuple<I8, I8, F32, I32>>;
using AvgPool2D_F8_Types =
::testing::Types<std::tuple<F8, F8, F32, I32>, std::tuple<F8, F8, F32, I32>>;
template <typename TType>
class AvgPool2D_F32 : public TestAvgPool2dFwd<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP32)
{
GTEST_SKIP() << "Skipping AvgPool2D_F32 tests because CK_ENABLE_FP32 is "
"not enabled";
}
}
};
template <typename TType>
class AvgPool2D_F16 : public TestAvgPool2dFwd<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP16)
{
GTEST_SKIP() << "Skipping AvgPool2D_F16 tests because CK_ENABLE_FP16 is "
"not enabled";
}
}
};
template <typename TType>
class AvgPool2D_BF16 : public TestAvgPool2dFwd<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_BF16)
{
GTEST_SKIP() << "Skipping AvgPool2D_BF16 tests because CK_ENABLE_BF16 is "
"not enabled";
}
}
};
template <typename TType>
class AvgPool2D_I8 : public TestAvgPool2dFwd<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_INT8)
{
GTEST_SKIP() << "Skipping AvgPool2D_I8 tests because CK_ENABLE_INT8 is "
"not enabled";
}
}
};
template <typename TType>
class AvgPool2D_F8 : public TestAvgPool2dFwd<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP8)
{
GTEST_SKIP() << "Skipping AvgPool2D_F8 tests because CK_ENABLE_FP8 is "
"not enabled";
}
}
};
TYPED_TEST_SUITE(AvgPool2D_F32, AvgPool2D_F32_Types);
TYPED_TEST_SUITE(AvgPool2D_F16, AvgPool2D_F16_Types);
TYPED_TEST_SUITE(AvgPool2D_BF16, AvgPool2D_BF16_Types);
TYPED_TEST_SUITE(AvgPool2D_I8, AvgPool2D_I8_Types);
TYPED_TEST_SUITE(AvgPool2D_F8, AvgPool2D_F8_Types);
TYPED_TEST(AvgPool2D_F32, AvgPool2D_I8_Test) { this->Run(); }
TYPED_TEST(AvgPool2D_F16, AvgPool2D_F16_Test) { this->Run(); }
TYPED_TEST(AvgPool2D_BF16, AvgPool2D_BF16_Test) { this->Run(); }
TYPED_TEST(AvgPool2D_I8, AvgPool2D_I8_Test) { this->Run(); }
TYPED_TEST(AvgPool2D_F8, AvgPool2D_F8_Test) { this->Run(); }
test/pool/test_avg_pool3d_fwd.cpp
View file @ 09d4c3a4
// 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 "gtest/gtest.h"
#include "profiler/profile_pool3d_fwd_impl.hpp"
......@@ -16,10 +16,19 @@ class TestAvgPool3dFwd : public ::testing::Test
std::vector<PoolingParam> params;
ck::profiler::PoolFwdInputParams in_params_avg_pool{true, 2, false, false, false, 1};
void Run()
{
for(auto param : params)
{
ck::profiler::PoolFwdKernelParams kernel_params{param.length_,
param.window_spatial_lengths_,
param.window_strides_,
param.window_dilations_,
param.input_left_pads_,
param.input_right_pads_};
bool success =
ck::profiler::profile_pool3d_fwd_impl<InDataType,
OutDataType,
......@@ -29,26 +38,18 @@ class TestAvgPool3dFwd : public ::testing::Test
ck::tensor_layout::convolution::NDHWC,
ck::ReduceTensorOp::AVG,
false,
false>(true,
2,
false,
false,
param.length_,
param.window_spatial_lengths_,
param.window_strides_,
param.window_dilations_,
param.input_left_pads_,
param.input_right_pads_);
false>(in_params_avg_pool, kernel_params);
EXPECT_TRUE(success);
}
}
};
#ifdef CK_ENABLE_FP16
using KernelTypes =
::testing::Types<std::tuple<F16, F16, F32, I32>, std::tuple<F32, F32, F32, I32>>;
#else
using KernelTypes = ::testing::Types<std::tuple<F32, F32, F32, I32>>;
#endif
using KernelTypes = ::testing::Types<std::tuple<I8, I8, I32, I32>,
std::tuple<F8, F8, F32, I32>,
std::tuple<F16, F16, F32, I32>,
std::tuple<BF16, BF16, F32, I32>,
std::tuple<F32, F32, F32, I32>>;
TYPED_TEST_SUITE(TestAvgPool3dFwd, KernelTypes);
TYPED_TEST(TestAvgPool3dFwd, Test_Pool)
{
......
test/pool/test_max_pool2d_bwd.cpp 0 → 100644
View file @ 09d4c3a4
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "profiler/profile_max_pool2d_bwd_impl.hpp"
#include "test_pool_fwd_common.hpp"
template <typename T>
class MaxPool2dBWDTest : public ::testing::Test
{
protected:
using DOutDataType = std::tuple_element_t<0, T>;
using DInDataType = std::tuple_element_t<1, T>;
using IndexDataType = std::tuple_element_t<2, T>;
using InDataType = DInDataType;
using OutDataType = DOutDataType;
static std::vector<PoolingParam> params;
void Run()
{
for(auto param : this->params)
{
bool success =
ck::profiler::profile_max_pool2d_bwd_impl<InDataType,
OutDataType,
IndexDataType,
DOutDataType,
DInDataType,
false>(true,
2,
false,
false,
param.length_,
param.window_spatial_lengths_,
param.window_strides_,
param.window_dilations_,
param.input_left_pads_,
param.input_right_pads_);
EXPECT_TRUE(success);
}
}
};
template <typename T>
std::vector<PoolingParam> MaxPool2dBWDTest<T>::params = {
{{1, 1, 1, 1}, {1, 1}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
{{2, 16, 64, 64}, {64, 64}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
{{2, 16, 64, 64}, {4, 4}, {4, 4}, {2, 2}, {0, 0}, {0, 0}},
{{2, 32, 30, 30}, {2, 2}, {2, 2}, {1, 1}, {1, 1}, {1, 1}},
{{2, 2, 30, 30}, {2, 2}, {2, 2}, {1, 1}, {1, 1}, {1, 1}}};
using Max_Pool_2D_f32_types = ::testing::Types<std::tuple<F32, F32, I32>>;
using Max_Pool_2D_int8_types = ::testing::Types<std::tuple<I8, I8, I32>>;
using Max_Pool_2D_f16_types = ::testing::Types<std::tuple<F16, F16, I32>>;
using Max_Pool_2D_bf16_types = ::testing::Types<std::tuple<BF16, BF16, I32>>;
using Max_Pool_2D_f8_types = ::testing::Types<std::tuple<F8, F8, I32>>;
template <typename TType>
class MaxPool2D_f32 : public MaxPool2dBWDTest<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP32)
{
GTEST_SKIP() << "Skipping MaxPool2D_f32 tests because CK_ENABLE_FP32 is not enabled";
}
}
};
template <typename TType>
class MaxPool2D_int8 : public MaxPool2dBWDTest<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_INT8)
{
GTEST_SKIP() << "Skipping MaxPool2D_int8 tests because CK_ENABLE_INT8 is not enabled";
}
}
};
template <typename TType>
class MaxPool2D_f16 : public MaxPool2dBWDTest<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP16)
{
GTEST_SKIP() << "Skipping MaxPool2D_f16 because CK_ENABLE_FP16 is not enabled";
}
}
};
template <typename TType>
class MaxPool2D_bf16 : public MaxPool2dBWDTest<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_BF16)
{
GTEST_SKIP() << "Skipping MaxPool2D_bf16 tests because CK_ENABLE_BF16 is not enabled";
}
}
};
template <typename TType>
class MaxPool2D_f8 : public MaxPool2dBWDTest<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP8)
{
GTEST_SKIP() << "Skipping MaxPool2D_f8 tests because CK_ENABLE_FP8 is not enabled";
}
}
};
TYPED_TEST_SUITE(MaxPool2D_f32, Max_Pool_2D_f32_types);
TYPED_TEST_SUITE(MaxPool2D_int8, Max_Pool_2D_int8_types);
TYPED_TEST_SUITE(MaxPool2D_f16, Max_Pool_2D_f16_types);
TYPED_TEST_SUITE(MaxPool2D_bf16, Max_Pool_2D_bf16_types);
TYPED_TEST_SUITE(MaxPool2D_f8, Max_Pool_2D_f8_types);
TYPED_TEST(MaxPool2D_f32, MaxPool2DTest_f32) { this->Run(); }
TYPED_TEST(MaxPool2D_int8, MaxPool2DTest_int8) { this->Run(); }
TYPED_TEST(MaxPool2D_f16, MaxPool2DTest_f16) { this->Run(); }
TYPED_TEST(MaxPool2D_bf16, MaxPool2DTest_bf16) { this->Run(); }
TYPED_TEST(MaxPool2D_f8, MaxPool2DTest_f8) { this->Run(); }
test/pool/test_max_pool2d_fwd.cpp 0 → 100644
View file @ 09d4c3a4
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "gtest/gtest.h"
#include "profiler/profile_pool2d_fwd_impl.hpp"
#include "test_pool_fwd_common.hpp"
template <typename Tuple>
class TestMaxPool2dFwd : public ::testing::Test
{
protected:
using InDataType = std::tuple_element_t<0, Tuple>;
using OutDataType = std::tuple_element_t<1, Tuple>;
using ComputeDataType = std::tuple_element_t<2, Tuple>;
using IndexDataType = std::tuple_element_t<3, Tuple>;
static constexpr bool ReturnIndex = std::tuple_element_t<4, Tuple>::value;
static std::vector<PoolingParam> params;
void Run()
{
for(auto param : params)
{
// max pool
bool success =
ck::profiler::profile_pool2d_fwd_impl<InDataType,
OutDataType,
ComputeDataType,
IndexDataType,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::NHWC,
ck::ReduceTensorOp::MAX,
false,
ReturnIndex>(true,
2,
false,
false,
param.length_,
param.window_spatial_lengths_,
param.window_strides_,
param.window_dilations_,
param.input_left_pads_,
param.input_right_pads_);
EXPECT_TRUE(success);
}
}
};
template <typename T>
std::vector<PoolingParam> TestMaxPool2dFwd<T>::params = {
{{{1, 1, 1, 1}, {1, 1}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
{{2, 16, 64, 64}, {64, 64}, {1, 1}, {1, 1}, {0, 0}, {0, 0}},
{{2, 16, 64, 64}, {4, 4}, {4, 4}, {2, 2}, {0, 0}, {0, 0}},
{{2, 32, 30, 30}, {2, 2}, {2, 2}, {1, 1}, {1, 1}, {1, 1}}}};
using true_t = std::integral_constant<bool, true>;
using false_t = std::integral_constant<bool, false>;
using MaxPool2D_F32_Types = ::testing::Types<std::tuple<F32, F32, F32, I32, true_t>,
std::tuple<F32, F32, F32, I32, false_t>>;
using MaxPool2D_F16_Types = ::testing::Types<std::tuple<F16, F16, F32, I32, true_t>,
std::tuple<F16, F16, F32, I32, false_t>>;
using MaxPool2D_BF16_Types = ::testing::Types<std::tuple<I8, I8, F32, I32, true_t>,
std::tuple<BF16, BF16, F32, I32, false_t>>;
using MaxPool2D_I8_Types =
::testing::Types<std::tuple<I8, I8, F32, I32, true_t>, std::tuple<I8, I8, F32, I32, false_t>>;
using MaxPool2D_F8_Types =
::testing::Types<std::tuple<F8, F8, F32, I32, true_t>, std::tuple<F8, F8, F32, I32, false_t>>;
template <typename TType>
class MaxPool2D_F32 : public TestMaxPool2dFwd<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP32)
{
GTEST_SKIP() << "Skipping MaxPool2D_F32 tests because CK_ENABLE_FP32 is "
"not enabled";
}
}
};
template <typename TType>
class MaxPool2D_F16 : public TestMaxPool2dFwd<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP16)
{
GTEST_SKIP() << "Skipping MaxPool2D_F16 tests because CK_ENABLE_FP16 is "
"not enabled";
}
}
};
template <typename TType>
class MaxPool2D_BF16 : public TestMaxPool2dFwd<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_BF16)
{
GTEST_SKIP() << "Skipping MaxPool2D_BF16 tests because CK_ENABLE_BF16 is "
"not enabled";
}
}
};
template <typename TType>
class MaxPool2D_I8 : public TestMaxPool2dFwd<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_INT8)
{
GTEST_SKIP() << "Skipping MaxPool2D_I8 tests because CK_ENABLE_INT8 is "
"not enabled";
}
}
};
template <typename TType>
class MaxPool2D_F8 : public TestMaxPool2dFwd<TType>
{
protected:
void SetUp() override
{
if(!CK_ENABLE_FP8)
{
GTEST_SKIP() << "Skipping MaxPool2D_F8 tests because CK_ENABLE_FP8 is "
"not enabled";
}
}
};
TYPED_TEST_SUITE(MaxPool2D_F32, MaxPool2D_F32_Types);
TYPED_TEST_SUITE(MaxPool2D_F16, MaxPool2D_F16_Types);
TYPED_TEST_SUITE(MaxPool2D_BF16, MaxPool2D_BF16_Types);
TYPED_TEST_SUITE(MaxPool2D_I8, MaxPool2D_I8_Types);
TYPED_TEST_SUITE(MaxPool2D_F8, MaxPool2D_F8_Types);
TYPED_TEST(MaxPool2D_F32, MaxPool2D_I8_Test) { this->Run(); }
TYPED_TEST(MaxPool2D_F16, MaxPool2D_F16_Test) { this->Run(); }
TYPED_TEST(MaxPool2D_BF16, MaxPool2D_BF16_Test) { this->Run(); }
TYPED_TEST(MaxPool2D_I8, MaxPool2D_I8_Test) { this->Run(); }
TYPED_TEST(MaxPool2D_F8, MaxPool2D_F8_Test) { this->Run(); }
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