Skip to content

GitLab

"...resnet50_tensorflow.git" did not exist on "503cc99cffb04b688966da832c5ff5bb2eb6f4a9"
Unverified Commit aebd211c authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

External Interface (#304) 

* add client example

* clean

* clean

* reorg

* clean up profiler

* reorg

* clea

* fix profiler

* function for getinstances

* update client example

* update client example

* update client example

* update

* update example

* update Jenkins file

* update cmake

* update Jenkins
parent b653c5eb
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 442 additions and 918 deletions
profiler/src/profile_batched_gemm.cpp
View file @ aebd211c
...@@ -15,10 +15,6 @@ enum struct GemmMatrixLayout ...@@ -15,10 +15,6 @@ enum struct GemmMatrixLayout
MK_NK_MN, // 1 MK_NK_MN, // 1
KM_KN_MN, // 2 KM_KN_MN, // 2
KM_NK_MN, // 3 KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
}; };
enum struct GemmDataType enum struct GemmDataType
...@@ -31,7 +27,7 @@ enum struct GemmDataType ...@@ -31,7 +27,7 @@ enum struct GemmDataType
int profile_batched_gemm(int argc, char* argv[]) int profile_batched_gemm(int argc, char* argv[])
{ {
if(!(argc == 15)) if(argc != 15)
{ {
printf("arg1: tensor operation (batched_gemm: Batched GEMM)\n"); printf("arg1: tensor operation (batched_gemm: Batched GEMM)\n");
printf("arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n"); printf("arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n");
...@@ -64,330 +60,117 @@ int profile_batched_gemm(int argc, char* argv[]) ...@@ -64,330 +60,117 @@ int profile_batched_gemm(int argc, char* argv[])
const int BatchCount = std::stoi(argv[14]); const int BatchCount = std::stoi(argv[14]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN) using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto c_type,
auto a_layout,
auto b_layout,
auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using CDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CLayout = decltype(c_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
bool pass = ck::profiler::
profile_batched_gemm_impl<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideC < 0) ? DefaultStrideC : StrideC,
BatchCount);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{ {
ck::profiler::profile_batched_gemm_impl<ck::half_t, return profile(F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
ck::half_t,
ck::half_t,
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) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN) else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{ {
ck::profiler::profile_batched_gemm_impl<ck::half_t, return profile(F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
ck::half_t,
ck::half_t,
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) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN) else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{ {
ck::profiler::profile_batched_gemm_impl<ck::half_t, return profile(F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
ck::half_t,
ck::half_t,
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,
BatchCount);
} }
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN) else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{ {
ck::profiler::profile_batched_gemm_impl<ck::half_t, return profile(F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
ck::half_t,
ck::half_t,
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) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN) else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{ {
ck::profiler::profile_batched_gemm_impl<ck::bhalf_t, return profile(F16{}, F16{}, F16{}, Row{}, Row{}, Row{});
ck::bhalf_t,
ck::bhalf_t,
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) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN) else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{ {
ck::profiler::profile_batched_gemm_impl<ck::bhalf_t, return profile(F16{}, F16{}, F16{}, Row{}, Col{}, Row{});
ck::bhalf_t,
ck::bhalf_t,
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) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN) else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{ {
ck::profiler::profile_batched_gemm_impl<ck::bhalf_t, return profile(F16{}, F16{}, F16{}, Col{}, Row{}, Row{});
ck::bhalf_t,
ck::bhalf_t,
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,
BatchCount);
} }
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN) else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{ {
ck::profiler::profile_batched_gemm_impl<ck::bhalf_t, return profile(F16{}, F16{}, F16{}, Col{}, Col{}, Row{});
ck::bhalf_t,
ck::bhalf_t,
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) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN) else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{ {
ck::profiler::profile_batched_gemm_impl<float, return profile(BF16{}, BF16{}, BF16{}, Row{}, Row{}, Row{});
float,
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) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN) else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{ {
ck::profiler::profile_batched_gemm_impl<float, return profile(BF16{}, BF16{}, BF16{}, Row{}, Col{}, Row{});
float,
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) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN) else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
{ {
ck::profiler::profile_batched_gemm_impl<float, return profile(BF16{}, BF16{}, BF16{}, Col{}, Row{}, Row{});
float,
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,
BatchCount);
} }
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN) else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
{ {
ck::profiler::profile_batched_gemm_impl<float, return profile(BF16{}, BF16{}, BF16{}, Col{}, Col{}, Row{});
float,
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) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN) else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
{ {
ck::profiler::profile_batched_gemm_impl<int8_t, return profile(INT8{}, INT8{}, INT8{}, Row{}, Row{}, Row{});
int8_t,
int8_t,
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) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN) else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
{ {
ck::profiler::profile_batched_gemm_impl<int8_t, return profile(INT8{}, INT8{}, INT8{}, Row{}, Col{}, Row{});
int8_t,
int8_t,
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) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN) else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
{ {
ck::profiler::profile_batched_gemm_impl<int8_t, return profile(INT8{}, INT8{}, INT8{}, Col{}, Row{}, Row{});
int8_t,
int8_t,
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,
BatchCount);
} }
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN) else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
{ {
ck::profiler::profile_batched_gemm_impl<int8_t, return profile(INT8{}, INT8{}, INT8{}, Col{}, Col{}, Row{});
int8_t,
int8_t,
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) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
BatchCount);
} }
else else
{ {
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented"); std::cout << "this data_type & layout is not implemented" << std::endl;
}
return 0; return 1;
}
} }
profiler/src/profile_convnd_fwd.cpp
View file @ aebd211c
...@@ -10,11 +10,10 @@ ...@@ -10,11 +10,10 @@
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/conv_util.hpp" #include "ck/library/utility/conv_util.hpp"
#include "ck/library/utility/fill.hpp" #include "ck/library/utility/fill.hpp"
#include "profiler/include/profile_convnd_fwd.hpp"
namespace { namespace {
enum struct ConvDataType enum struct ConvDataType
...@@ -304,7 +303,7 @@ void profile_convnd_instances(ConvDataType data_type, ...@@ -304,7 +303,7 @@ void profile_convnd_instances(ConvDataType data_type,
} // namespace } // namespace
int ck::profiler::profile_convnd_fwd(int argc, char* argv[]) int profile_convnd_fwd(int argc, char* argv[])
{ {
using namespace ck::utils::conv; using namespace ck::utils::conv;
......
profiler/src/profile_gemm.cpp
View file @ aebd211c
...@@ -14,10 +14,6 @@ enum struct GemmMatrixLayout ...@@ -14,10 +14,6 @@ enum struct GemmMatrixLayout
MK_NK_MN, // 1 MK_NK_MN, // 1
KM_KN_MN, // 2 KM_KN_MN, // 2
KM_NK_MN, // 3 KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
}; };
enum struct GemmDataType enum struct GemmDataType
...@@ -30,7 +26,7 @@ enum struct GemmDataType ...@@ -30,7 +26,7 @@ enum struct GemmDataType
int profile_gemm(int argc, char* argv[]) int profile_gemm(int argc, char* argv[])
{ {
if(!(argc == 14 || argc == 15)) if(argc != 14)
{ {
printf("arg1: tensor operation (gemm: GEMM)\n"); printf("arg1: tensor operation (gemm: GEMM)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"); printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
...@@ -41,9 +37,8 @@ int profile_gemm(int argc, char* argv[]) ...@@ -41,9 +37,8 @@ int profile_gemm(int argc, char* argv[])
printf("arg4: verification (0: no; 1: yes)\n"); printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n"); printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n"); printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n"); printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n"); printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: split k into mulitiple batch\n");
exit(1); exit(1);
} }
...@@ -61,350 +56,125 @@ int profile_gemm(int argc, char* argv[]) ...@@ -61,350 +56,125 @@ int profile_gemm(int argc, char* argv[])
const int StrideA = std::stoi(argv[11]); const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]); const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]); const int StrideC = std::stoi(argv[13]);
int KBatch = 1;
if(argc == 15)
KBatch = std::stoi(argv[14]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN) using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
using INT32 = int32_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto c_type,
auto a_layout,
auto b_layout,
auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using CDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CLayout = decltype(c_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
bool pass =
ck::profiler::profile_gemm_impl<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideC < 0) ? DefaultStrideC : StrideC);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{ {
ck::profiler::profile_gemm_impl<ck::half_t, return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
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) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN) else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{ {
ck::profiler::profile_gemm_impl<ck::half_t, return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
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) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN) else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{ {
ck::profiler::profile_gemm_impl<ck::half_t, return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
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,
KBatch);
} }
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN) else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{ {
ck::profiler::profile_gemm_impl<ck::half_t, return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
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) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN) else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{ {
ck::profiler::profile_gemm_impl<float, return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
float,
float,
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) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN) else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{ {
ck::profiler::profile_gemm_impl<float, return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
float,
float,
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) ? K : StrideA,
(StrideB < 0) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN) else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{ {
ck::profiler::profile_gemm_impl<float, return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
float,
float,
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,
KBatch);
} }
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN) else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{ {
ck::profiler::profile_gemm_impl<float, return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
float,
float,
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) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN) else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
{ {
ck::profiler::profile_gemm_impl<int8_t, return profile(BF16{}, BF16{}, F32{}, BF16{}, Row{}, Row{}, Row{});
int8_t,
int8_t,
int32_t,
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) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN) else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
{ {
ck::profiler::profile_gemm_impl<int8_t, return profile(BF16{}, BF16{}, F32{}, BF16{}, Row{}, Col{}, Row{});
int8_t,
int8_t,
int32_t,
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) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN) else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
{ {
ck::profiler::profile_gemm_impl<int8_t, return profile(BF16{}, BF16{}, F32{}, BF16{}, Col{}, Row{}, Row{});
int8_t,
int8_t,
int32_t,
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,
KBatch);
} }
else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN) else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
{ {
ck::profiler::profile_gemm_impl<int8_t, return profile(BF16{}, BF16{}, F32{}, BF16{}, Col{}, Col{}, Row{});
int8_t,
int8_t,
int32_t,
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) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN) else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
{ {
ck::profiler::profile_gemm_impl<ck::bhalf_t, return profile(INT8{}, INT8{}, INT32{}, INT8{}, Row{}, Row{}, Row{});
ck::bhalf_t,
ck::bhalf_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) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN) else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
{ {
ck::profiler::profile_gemm_impl<ck::bhalf_t, return profile(INT8{}, INT8{}, INT32{}, INT8{}, Row{}, Col{}, Row{});
ck::bhalf_t,
ck::bhalf_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) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN) else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
{ {
ck::profiler::profile_gemm_impl<ck::bhalf_t, return profile(INT8{}, INT8{}, INT32{}, INT8{}, Col{}, Row{}, Row{});
ck::bhalf_t,
ck::bhalf_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,
KBatch);
} }
else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN) else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
{ {
ck::profiler::profile_gemm_impl<ck::bhalf_t, return profile(INT8{}, INT8{}, INT32{}, INT8{}, Col{}, Col{}, Row{});
ck::bhalf_t,
ck::bhalf_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) ? K : StrideB,
(StrideC < 0) ? N : StrideC,
KBatch);
} }
else else
{ {
throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented"); std::cout << "this data_type & layout is not implemented" << std::endl;
}
return 0; return 1;
}
} }
profiler/src/profile_gemm_add_add_fastgelu.cpp
View file @ aebd211c
...@@ -16,10 +16,6 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[]) ...@@ -16,10 +16,6 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
MK_NK_MN_MN_MN, // 1 MK_NK_MN_MN_MN, // 1
KM_KN_MN_MN_MN, // 2 KM_KN_MN_MN_MN, // 2
KM_NK_MN_MN_MN, // 3 KM_NK_MN_MN_MN, // 3
MK_KN_NM_MN_MN, // 4
MK_NK_NM_MN_MN, // 5
KM_KN_NM_MN_MN, // 6
KM_NK_NM_MN_MN, // 7
}; };
enum struct MatrixDataType enum struct MatrixDataType
...@@ -101,17 +97,17 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[]) ...@@ -101,17 +97,17 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M; const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M; const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
return ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType, bool pass = ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType,
BDataType, BDataType,
AccDataType, AccDataType,
D0DataType, D0DataType,
D1DataType, D1DataType,
EDataType, EDataType,
ALayout, ALayout,
BLayout, BLayout,
D0Layout, D0Layout,
D1Layout, D1Layout,
ELayout>( ELayout>(
do_verification, do_verification,
init_method, init_method,
do_log, do_log,
...@@ -124,6 +120,8 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[]) ...@@ -124,6 +120,8 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0, (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideD1 < 0) ? DefaultStrideD1 : StrideD1, (StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
(StrideE < 0) ? DefaultStrideE : StrideE); (StrideE < 0) ? DefaultStrideE : StrideE);
return pass ? 0 : 1;
}; };
if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN) if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
...@@ -149,6 +147,6 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[]) ...@@ -149,6 +147,6 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
{ {
std::cout << "this data_type & layout is not implemented" << std::endl; std::cout << "this data_type & layout is not implemented" << std::endl;
return 0; return 1;
} }
} }
profiler/src/profile_gemm_splitk.cpp 0 → 100644
View file @ aebd211c
// 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_splitk_impl.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
};
int profile_gemm_splitk(int argc, char* argv[])
{
if(argc != 15)
{
printf("arg1: tensor operation (gemm_splitk: Split-K GEMM)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\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=no, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: split k into mulitiple batch\n");
exit(1);
}
const auto data_type = static_cast<GemmDataType>(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 KBatch = std::stoi(argv[14]);
using F32 = float;
using F16 = ck::half_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto c_type,
auto a_layout,
auto b_layout,
auto c_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using CDataType = decltype(c_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CLayout = decltype(c_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_splitk_impl<ADataType,
BDataType,
AccDataType,
CDataType,
ALayout,
BLayout,
CLayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideC < 0) ? DefaultStrideC : StrideC,
KBatch);
return pass ? 0 : 1;
};
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
profiler/src/profiler.cpp
View file @ aebd211c
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <cstring> #include <cstring>
#include "profiler/include/profile_convnd_fwd.hpp"
int profile_gemm(int, char*[]); int profile_gemm(int, char*[]);
int profile_gemm_splitk(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_bias_add_reduce(int, char*[]); int profile_gemm_bias_add_reduce(int, char*[]);
int profile_gemm_add_add_fastgelu(int, char*[]);
int profile_gemm_reduce(int, char*[]);
int profile_batched_gemm(int, char*[]); int profile_batched_gemm(int, char*[]);
int profile_batched_gemm_reduce(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_convnd_fwd(int argc, char* argv[]);
int profile_convnd_bwd_data(int, char*[], int); int profile_convnd_bwd_data(int, char*[], int);
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_reduce(int, char*[]);
int profile_gemm_add_add_fastgelu(int, char*[]);
static void print_helper_message() static void print_helper_message()
{ {
// clang-format off // clang-format off
printf("arg1: tensor operation (gemm: GEMM\n" printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_bias_2d: GEMM+Bias(2D)\n" " gemm_splitk: Split-K GEMM\n"
" gemm_bias_relu: GEMM+Bias+ReLU\n" " gemm_bias_2d: GEMM+Bias(2D)\n"
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n" " gemm_bias_relu: GEMM+Bias+ReLU\n"
" gemm_reduce: GEMM+Reduce\n" " gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
" grouped_gemm: Grouped GEMM\n" " gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n"
" conv_fwd: ForwardConvolution\n" " gemm_reduce: GEMM+Reduce\n"
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n" " batched_gemm: Batched GEMM\n"
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n" " grouped_gemm: Grouped GEMM\n"
" conv1d_bwd_data: BackwardConvolution data 1 dim\n" " conv_fwd: ForwardConvolution\n"
" conv2d_bwd_data: BackwardConvolution data 2 dim\n" " conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
" conv3d_bwd_data: BackwardConvolution data 3 dim\n" " conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
" reduce: Reduce\n" " conv1d_bwd_data: BackwardConvolution data 1 dim\n"
" conv2d_bwd_weight: Backward Weight Convolution 2d\n" " conv2d_bwd_data: BackwardConvolution data 2 dim\n"
" gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n"); " conv3d_bwd_data: BackwardConvolution data 3 dim\n"
" conv2d_bwd_weight: Backward Weight Convolution 2d\n"
" reduce: Reduce\n");
// clang-format on // clang-format on
} }
...@@ -60,6 +58,10 @@ int main(int argc, char* argv[]) ...@@ -60,6 +58,10 @@ int main(int argc, char* argv[])
{ {
return profile_gemm(argc, argv); return profile_gemm(argc, argv);
} }
else if(strcmp(argv[1], "gemm_splitk") == 0)
{
return profile_gemm_splitk(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_2d") == 0) else if(strcmp(argv[1], "gemm_bias_2d") == 0)
{ {
return profile_gemm_bias_2d(argc, argv); return profile_gemm_bias_2d(argc, argv);
...@@ -94,7 +96,7 @@ int main(int argc, char* argv[]) ...@@ -94,7 +96,7 @@ int main(int argc, char* argv[])
} }
else if(strcmp(argv[1], "conv_fwd") == 0) else if(strcmp(argv[1], "conv_fwd") == 0)
{ {
return ck::profiler::profile_convnd_fwd(argc, argv); return profile_convnd_fwd(argc, argv);
} }
else if(strcmp(argv[1], "conv_fwd_bias_relu") == 0) else if(strcmp(argv[1], "conv_fwd_bias_relu") == 0)
{ {
......
test/batched_gemm/batched_gemm_util.hpp deleted 100644 → 0
View file @ b653c5eb
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef BATCHED_GEMM_UTILS_HPP
#define BATCHED_GEMM_UTILS_HPP
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
namespace ck {
namespace batched_gemm_util {
struct GemmParams
{
GemmParams()
: M(1024), N(1024), K(1024), StrideA(1024), StrideB(1024), StrideC(1024), alpha(1), beta(0)
{
}
ck::index_t M;
ck::index_t N;
ck::index_t K;
ck::index_t StrideA;
ck::index_t StrideB;
ck::index_t StrideC;
float alpha;
float beta;
};
template <typename BatchedGemmInstance,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunHostBatchedGemm(const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
auto ref_batched_gemm = BatchedGemmInstance{};
auto ref_invoker = ref_batched_gemm.MakeInvoker();
auto ref_argument =
ref_batched_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
template <typename DeviceGemmPtr,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunDeviceBatchedGemm(DeviceGemmPtr& batched_gemm_ptr,
const ck::batched_gemm_util::GemmParams& params,
const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpace());
DeviceMem b_g_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpace());
DeviceMem c_g_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpace());
a_g_m_k_device_buf.ToDevice(A.mData.data());
b_g_k_n_device_buf.ToDevice(B.mData.data());
const auto batch_count = A.mDesc.GetLengths()[0];
auto invoker_ptr = batched_gemm_ptr->MakeInvokerPointer();
auto argument_ptr = batched_gemm_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_g_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_g_m_n_device_buf.GetDeviceBuffer()),
params.M,
params.N,
params.K,
params.StrideA,
params.StrideB,
params.StrideC,
a_element_op,
b_element_op,
c_element_op,
batch_count);
if(!batched_gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
invoker_ptr->Run(argument_ptr.get());
c_g_m_n_device_buf.FromDevice(C.mData.data());
}
} // namespace batched_gemm_util
} // namespace ck
#endif
test/gemm/gemm_util.hpp
View file @ aebd211c
...@@ -214,6 +214,11 @@ struct TestGemm ...@@ -214,6 +214,11 @@ struct TestGemm
res = ck::utils::check_err(c_device.mData, c_host.mData); res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl; std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
} }
else if(std::is_same<CDataType, ck::bhalf_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, int8_t>::value) else if(std::is_same<CDataType, int8_t>::value)
{ {
res = ck::utils::check_err(c_device.mData, c_host.mData); res = ck::utils::check_err(c_device.mData, c_host.mData);
...@@ -234,121 +239,5 @@ struct TestGemm ...@@ -234,121 +239,5 @@ struct TestGemm
} }
}; };
template <typename DeviceGemmPtr_,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct TestGemmBF16
{
using BF16 = ck::bhalf_t;
auto PrepareGemmTensorBF16(const ck::gemm_util::GemmParams& params)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
// use fp32 host kernel to verify bf16 device kernel
Tensor<BF16> a_m_k_bf16(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<BF16> b_k_n_bf16(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<BF16> c_m_n_device_bf16(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> a_m_k_fp32(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<float> b_k_n_fp32(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<float> c_m_n_host_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> c_m_n_device_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
a_m_k_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
b_k_n_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
bf16_to_f32_(a_m_k_bf16, a_m_k_fp32);
bf16_to_f32_(b_k_n_bf16, b_k_n_fp32);
return std::make_tuple(a_m_k_bf16,
b_k_n_bf16,
c_m_n_device_bf16,
a_m_k_fp32,
b_k_n_fp32,
c_m_n_host_fp32,
c_m_n_device_fp32);
}
auto operator()(DeviceGemmPtr_& gemmPtr)
{
// Arrange
ck::gemm_util::GemmParams params;
params.M = 1024;
params.N = 1024;
params.K = 1024;
params.StrideA = 1024;
params.StrideB = 1024;
params.StrideC = 1024;
auto host_tensors = PrepareGemmTensorBF16(params);
const Tensor<BF16>& a_bf16 = std::get<0>(host_tensors);
const Tensor<BF16>& b_bf16 = std::get<1>(host_tensors);
Tensor<BF16>& c_device_bf16 = std::get<2>(host_tensors);
Tensor<float>& a_fp32 = std::get<3>(host_tensors);
Tensor<float>& b_fp32 = std::get<4>(host_tensors);
Tensor<float>& c_host_fp32 = std::get<5>(host_tensors);
Tensor<float>& c_device_fp32 = std::get<6>(host_tensors);
auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{};
auto c_element_op = CElementwiseOperation{};
// use fp32 host kernel to verify bf16 device kernel
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<float,
float,
float,
float,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
a_fp32, b_fp32, c_host_fp32, a_element_op, b_element_op, c_element_op);
// Act
ck::gemm_util::RunDeviceGEMM(gemmPtr,
params,
a_bf16,
b_bf16,
c_device_bf16,
a_element_op,
b_element_op,
c_element_op);
bf16_to_f32_(c_device_bf16, c_device_fp32);
// Assert
bool res = ck::utils::check_err(
c_device_fp32.mData, c_host_fp32.mData, "Error: incorrect results!", 1e-2f, 1e-3f);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res;
};
};
} // namespace gemm_util } // namespace gemm_util
} // namespace ck } // namespace ck
test/gemm/gemm_xdl_bf16.cpp
View file @ aebd211c
This diff is collapsed.
test/gemm/gemm_xdl_fp16.cpp
View file @ aebd211c
This diff is collapsed.
test/gemm/gemm_xdl_fp32.cpp
View file @ aebd211c
This diff is collapsed.
test/gemm_split_k/CMakeLists.txt
View file @ aebd211c
add_test_executable(test_gemm_split_k gemm_split_k.cpp) add_test_executable(test_gemm_split_k gemm_split_k.cpp)
target_link_libraries(test_gemm_split_k PRIVATE host_tensor) target_link_libraries(test_gemm_split_k PRIVATE host_tensor)
target_link_libraries(test_gemm_split_k PRIVATE device_gemm_instance) target_link_libraries(test_gemm_split_k PRIVATE device_gemm_splitk_instance)
test/gemm_split_k/gemm_split_k.cpp
View file @ aebd211c
This diff is collapsed.
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