Skip to content

GitLab

"backend/apps/webui/internal/migrations/001_initial_schema.py" did not exist on "f8f9f27ae8369c63def383957896d2dfadbf2c07"
Commit 7e8230da authored by Jun Liu's avatar Jun Liu
Browse files

Merge branch 'develop' into amd-develop

parents 56c72035 bd09b5c5
Show whitespace changes
Inline Side-by-side
Showing with 651 additions and 272 deletions
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp 0 → 100644
View file @ 7e8230da
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_bwd_data_wmma_gndhwk_gkzyxc_gndhwc_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
GNDHWK,
GKZYXC,
Empty_Tuple,
GNDHWC,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_wmma_i8_instances<3,
GNDHWK,
GKZYXC,
Empty_Tuple,
GNDHWC,
Empty_Tuple,
PassThrough,
ConvBwdDataDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp 0 → 100644
View file @ 7e8230da
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_f16_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_wmma_f16_instances<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
Empty_Tuple,
PassThrough,
ConvBwdData1x1S1P0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp 0 → 100644
View file @ 7e8230da
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_wmma_f16_instances<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
Empty_Tuple,
PassThrough,
ConvBwdDataDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp 0 → 100644
View file @ 7e8230da
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_i8_1x1s1p0_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_wmma_i8_instances<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
Empty_Tuple,
PassThrough,
ConvBwdData1x1S1P0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp 0 → 100644
View file @ 7e8230da
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_bwd_data_wmma_ndhwgk_gkzyxc_ndhwgc_i8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_wmma_i8_instances<3,
NDHWGK,
GKZYXC,
Empty_Tuple,
NDHWGC,
Empty_Tuple,
PassThrough,
ConvBwdDataDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/CMakeLists.txt
View file @ 7e8230da
add_instance_library(device_grouped_gemm_fixed_nk_instance
device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp
device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp
set(GROUPED_GEMM_FIXED_NK_INSTANCES)
device_grouped_gemm_xdl_fixed_nk_f16_f8_f16_mk_kn_mn_instance.cpp
device_grouped_gemm_xdl_fixed_nk_f16_f8_f16_mk_nk_mn_instance.cpp
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp)
endif()
device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp
)
if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp)
endif()
if((DTYPES MATCHES "int8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp)
list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp)
endif()
add_instance_library(device_grouped_gemm_fixed_nk_instance ${GROUPED_GEMM_FIXED_NK_INSTANCES})
profiler/README.md
View file @ 7e8230da
......@@ -50,21 +50,23 @@ Best Perf: 1.42509 ms, 102.988 TFlops, 234.086 GB/s
## Profile contraction kernels
```bash
#arg1: tensor operation (contraction_bilinear=CONTRACTION+Bilinear)
#arg2: data type (0: fp32; 1: f64)\n"
#arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
#arg2: data type (0: fp32; 1: f64; 2: f16; 3: bf16)
#arg3: compute data type (0: fp32; 1: f64; 2: f16; 3: bf16)
#arg4: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
# 1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
# 2: A[k0, k1, m0, m1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
# 3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1])
#arg4: verification (0: no; 1: yes)
#arg5: initialization (0: no init; 1: integer value; 2: decimal value)
#arg6: print tensor value (0: no; 1: yes)
#arg7: time kernel (0: no, 1: yes)
#arg8 and arg9: alpha and beta
#arg10 to 15: M0, M1, N0, N1, K0, K1
#arg16 to 31: Strides for A, B, D and E (skip for default)
################ op datatype layout verify init log time alpha beta M0 M1 N0 N1 K0 K1
./bin/ckProfiler contraction_bilinear 0 1 0 0 0 1 1.0 1.0 128 128 128 128 128 128
#arg5: verification (0: no; 1: yes)
#arg6: initialization (0: no init; 1: integer value; 2: decimal value)
#arg7: print tensor value (0: no; 1: yes)
#arg8: time kernel (0: no, 1: yes)
#arg9: alpha
#arg10: beta
#arg11 to 16: M0, M1, N0, N1, K0, K1
#arg17 to 32: Strides for A, B, D and E (skip for default)
################ op datatype compute_datatype layout verify init log time alpha beta M0 M1 N0 N1 K0 K1
./bin/ckProfiler contraction_bilinear 0 0 1 0 0 0 1 1.0 1.0 128 128 128 128 128 128
```
Result (MI100)
......
profiler/include/profiler/profile_contraction_impl.hpp
View file @ 7e8230da
......@@ -31,10 +31,14 @@ namespace profiler {
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using Scale = ck::tensor_operation::element_wise::Scale;
using F32 = float;
using F64 = double;
template <typename ALayout,
typename BLayout,
typename CDELayout,
typename DataType,
typename ComputeDataType,
typename DTupleDataType,
typename CDElementOp>
int profile_contraction_impl(ck::index_t do_verification,
......@@ -45,10 +49,10 @@ int profile_contraction_impl(ck::index_t do_verification,
const std::vector<ck::index_t>& M,
const std::vector<ck::index_t>& N,
const std::vector<ck::index_t>& K,
const std::vector<ck::index_t>& StridesA,
const std::vector<ck::index_t>& StridesB,
const std::vector<ck::index_t>& StridesE,
const std::vector<ck::index_t>& StridesD)
const std::vector<ck::index_t>& StridesA, // [M0, M1, K0, K1]
const std::vector<ck::index_t>& StridesB, // [N0, N1, K0, K1]
const std::vector<ck::index_t>& StridesE, // [M0, M1, N0, N1]
const std::vector<ck::index_t>& StridesD) // [M0, M1, N0, N1]
{
bool pass = true;
......@@ -63,13 +67,13 @@ int profile_contraction_impl(ck::index_t do_verification,
};
Tensor<DataType> a_m_k(f_host_tensor_descriptor(M, K, StridesA));
Tensor<DataType> b_k_n(f_host_tensor_descriptor(K, N, StridesB));
Tensor<DataType> b_n_k(f_host_tensor_descriptor(N, K, StridesB));
Tensor<DataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StridesE));
Tensor<DataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StridesE));
Tensor<DataType> d_m_n(f_host_tensor_descriptor(M, N, StridesD));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "b_n_k: " << b_n_k.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
......@@ -78,12 +82,12 @@ int profile_contraction_impl(ck::index_t do_verification,
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
b_n_k.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<DataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5});
b_n_k.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5});
}
......@@ -91,12 +95,12 @@ int profile_contraction_impl(ck::index_t do_verification,
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
DeviceMem a_device_buf(sizeof(DataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(DataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(DataType) * b_n_k.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(DataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DataType) * d_m_n.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
b_device_buf.ToDevice(b_n_k.mData.data());
e_device_buf.SetZero();
d_device_buf.ToDevice(d_m_n.mData.data());
......@@ -118,7 +122,8 @@ int profile_contraction_impl(ck::index_t do_verification,
DataType,
AElementOp,
BElementOp,
CDElementOp>;
CDElementOp,
ComputeDataType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......@@ -126,6 +131,9 @@ int profile_contraction_impl(ck::index_t do_verification,
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
using AccDataType =
typename std::conditional<std::is_same<ComputeDataType, F64>::value, F64, F32>::type;
// Run reference op
if(do_verification)
{
......@@ -136,7 +144,8 @@ int profile_contraction_impl(ck::index_t do_verification,
DataType,
DataType,
DataType,
DataType,
AccDataType,
ComputeDataType,
AElementOp,
BElementOp>;
......@@ -146,7 +155,7 @@ int profile_contraction_impl(ck::index_t do_verification,
Tensor<DataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StridesE));
auto ref_argument =
ref_op.MakeArgument(a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op);
ref_op.MakeArgument(a_m_k, b_n_k, c_m_n_host_result, a_element_op, b_element_op);
ref_invoker.Run(ref_argument);
......@@ -272,8 +281,29 @@ int profile_contraction_impl(ck::index_t do_verification,
{
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
float threshold =
static_cast<DataType>(nelems_k) * std::numeric_limits<DataType>::epsilon();
// Both the kernel and the reference use `AccDataType`, so an absolute error of both
// of them is bounded by `nelems_k * std::numeric_limits<AccDataType>::epsilon()`.
// Comparing one to another can result in an absolute error as high as twice that
// value.
double threshold = 2 * nelems_k * std::numeric_limits<AccDataType>::epsilon();
// Handle the possible casting error of either AccDataType -> DataType or
// DataType -> ComputeDataType.
// TODO: Add a generic solution for calculating thresholds in CK.
if constexpr(ck::is_same_v<DataType, ck::bhalf_t> ||
ck::is_same_v<ComputeDataType, ck::bhalf_t>)
{
const double epsilon = std::pow(2, -7);
// Maximum relative casting error when rounding to zero.
threshold += epsilon * 2;
}
else if constexpr(ck::is_same_v<DataType, ck::half_t> ||
ck::is_same_v<ComputeDataType, ck::half_t>)
{
const double epsilon = std::pow(2, -10);
// Maximum relative casting error when rounding to zero.
threshold += epsilon * 2;
}
pass = pass & ck::utils::check_err(e_m_n_device_result,
e_m_n_host_result,
"Error: incorrect results!",
......@@ -283,7 +313,7 @@ int profile_contraction_impl(ck::index_t do_verification,
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_n_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", e_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", e_m_n_device_result.mData, ",")
......
profiler/include/profiler/profile_contraction_utils.hpp
View file @ 7e8230da
......@@ -25,6 +25,16 @@ enum struct ContractionDataType
{
F32_F32_F32_F32, // 0
F64_F64_F64_F64, // 1
F16_F16_F16_F16, // 2
BF16_BF16_BF16_BF16, // 3
};
enum struct ContractionComputeDataType
{
F32 = 0,
F64,
F16,
BF16,
};
inline void collect_index_params(char* argv[],
......
profiler/src/profile_contraction_bilinear.cpp
View file @ 7e8230da
......@@ -17,8 +17,9 @@
static void print_helper_msg()
{
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp32; 1: f64)\n"
<< "arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
<< "arg2: data type (0: fp32; 1: f64; 2: f16; 3: bf16)\n"
<< "arg3: compute data type (0: fp32; 1: f64; 2: f16; 3: bf16)\n"
<< "arg4: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
......@@ -26,40 +27,42 @@ static void print_helper_msg()
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1])\n"
<< "arg4: verification (0: no; 1: yes)\n"
<< "arg5: initialization (0: no init; 1: integer value; 2: decimal "
<< "arg5: verification (0: no; 1: yes)\n"
<< "arg6: initialization (0: no init; 1: integer value; 2: decimal "
<< "value)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0: no, 1: yes)\n"
<< "arg8 and arg9: alpha and beta\n"
<< "arg10 to 15: M0, M1, N0, N1, K0, K1\n"
<< "arg16 to 31: Strides for A, B, D and E (skip for default)\n"
<< "arg7: print tensor value (0: no; 1: yes)\n"
<< "arg8: time kernel (0: no, 1: yes)\n"
<< "arg9: alpha\n"
<< "arg10: beta\n"
<< "arg11 to 16: M0, M1, N0, N1, K0, K1\n"
<< "arg17 to 32: Strides for A, B, D and E (skip for default)\n"
<< std::endl;
}
int profile_contraction_bilinear(int argc, char* argv[])
{
const bool default_strides = argc == 16;
const bool default_strides = argc == 17;
if(argc != 32 && argc != 16)
if(argc != 33 && argc != 17)
{
print_helper_msg();
exit(1);
}
const auto data_type = static_cast<ContractionDataType>(std::stoi(argv[2]));
const auto layout = static_cast<ContractionMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const ck::index_t init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const float alpha = std::stof(argv[8]);
const float beta = std::stof(argv[9]);
const auto compute_data_type = static_cast<ContractionComputeDataType>(std::stoi(argv[3]));
const auto layout = static_cast<ContractionMatrixLayout>(std::stoi(argv[4]));
const bool do_verification = std::stoi(argv[5]);
const ck::index_t init_method = std::stoi(argv[6]);
const bool do_log = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[8]);
const float alpha = std::stof(argv[9]);
const float beta = std::stof(argv[10]);
std::vector<ck::index_t> M;
std::vector<ck::index_t> N;
std::vector<ck::index_t> K;
const ck::index_t dims_arg_num = 10;
const ck::index_t dims_arg_num = 11;
collect_index_params(argv, M, dims_arg_num, 2);
collect_index_params(argv, N, dims_arg_num + 2, 2);
collect_index_params(argv, K, dims_arg_num + 4, 2);
......@@ -76,20 +79,24 @@ int profile_contraction_bilinear(int argc, char* argv[])
collect_index_params(argv, StridesD, dims_arg_num + 18, 4);
}
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using F64 = double;
auto profile = [&](auto a_layout, auto b_layout, auto cde_layout, auto type) {
auto profile =
[&](auto a_layout, auto b_layout, auto cde_layout, auto type, auto compute_type) {
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CDELayout = decltype(cde_layout);
using DataType = decltype(type);
using ComputeDataType = decltype(compute_type);
if(default_strides)
{
assign_default_strides(a_layout, StridesA, {M[0], M[1], K[0], K[1]});
assign_default_strides(b_layout, StridesB, {K[0], K[1], N[0], N[1]});
assign_default_strides(b_layout, StridesB, {N[0], N[1], K[0], K[1]});
assign_default_strides(cde_layout, StridesE, {M[0], M[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesD, {M[0], M[1], N[0], N[1]});
}
......@@ -97,6 +104,7 @@ int profile_contraction_bilinear(int argc, char* argv[])
BLayout,
CDELayout,
DataType,
ComputeDataType,
ck::Tuple<DataType>,
Bilinear>(do_verification,
init_method,
......@@ -114,52 +122,87 @@ int profile_contraction_bilinear(int argc, char* argv[])
return pass;
};
if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
auto run_profile_for_datatype = [&](auto type, auto compute_type) {
if(layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
return profile(Row{}, Row{}, Row{}, F32{});
return profile(Row{}, Row{}, Row{}, type, compute_type);
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
else if(layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
return profile(Row{}, Col{}, Row{}, F32{});
return profile(Row{}, Col{}, Row{}, type, compute_type);
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
else if(layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
return profile(Col{}, Row{}, Row{}, F32{});
return profile(Col{}, Row{}, Row{}, type, compute_type);
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
else if(layout == ContractionMatrixLayout::KM_NK_MN_MN)
{
return profile(Col{}, Col{}, Row{}, F32{});
return profile(Col{}, Col{}, Row{}, type, compute_type);
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
return false;
};
if(data_type == ContractionDataType::F32_F32_F32_F32)
{
return profile(Row{}, Row{}, Row{}, F64{});
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F32{}, F32{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
else if(compute_data_type == ContractionComputeDataType::F16)
{
return profile(Row{}, Col{}, Row{}, F64{});
return run_profile_for_datatype(F32{}, F16{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
else if(compute_data_type == ContractionComputeDataType::BF16)
{
return profile(Col{}, Row{}, Row{}, F64{});
return run_profile_for_datatype(F32{}, BF16{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
else
{
return profile(Col{}, Col{}, Row{}, F64{});
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
else if(data_type == ContractionDataType::F64_F64_F64_F64)
{
if(compute_data_type == ContractionComputeDataType::F64)
{
return run_profile_for_datatype(F64{}, F64{});
}
else if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F64{}, F32{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
else if(data_type == ContractionDataType::F16_F16_F16_F16)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F16{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
else if(data_type == ContractionDataType::BF16_BF16_BF16_BF16)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(BF16{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
return 1;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_contraction_bilinear);
profiler/src/profile_contraction_scale.cpp
View file @ 7e8230da
......@@ -17,8 +17,9 @@
static void print_helper_msg()
{
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp32; 1: f64)\n"
<< "arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
<< "arg2: data type (0: fp32; 1: f64; 2: f16; 3: bf16)\n"
<< "arg3: compute data type (0: fp32; 1: f64; 2: f16; 3: bf16)\n"
<< "arg4: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
......@@ -26,39 +27,40 @@ static void print_helper_msg()
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1])\n"
<< "arg4: verification (0: no; 1: yes)\n"
<< "arg5: initialization (0: no init; 1: integer value; 2: decimal "
<< "arg5: verification (0: no; 1: yes)\n"
<< "arg6: initialization (0: no init; 1: integer value; 2: decimal "
<< "value)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0: no, 1: yes)\n"
<< "arg8: alpha\n"
<< "arg9 to 14: M0, M1, N0, N1, K0, K1\n"
<< "arg15 to 30: Strides for A, B, D and E (skip for default)\n"
<< "arg7: print tensor value (0: no; 1: yes)\n"
<< "arg8: time kernel (0: no, 1: yes)\n"
<< "arg9: alpha\n"
<< "arg10 to 15: M0, M1, N0, N1, K0, K1\n"
<< "arg16 to 31: Strides for A, B, D and E (skip for default)\n"
<< std::endl;
}
int profile_contraction_scale(int argc, char* argv[])
{
const bool default_strides = argc == 15;
const bool default_strides = argc == 16;
if(argc != 31 && argc != 15)
if(argc != 32 && argc != 16)
{
print_helper_msg();
exit(1);
}
const auto data_type = static_cast<ContractionDataType>(std::stoi(argv[2]));
const auto layout = static_cast<ContractionMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const ck::index_t init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const float alpha = std::stof(argv[8]);
const auto compute_data_type = static_cast<ContractionComputeDataType>(std::stoi(argv[3]));
const auto layout = static_cast<ContractionMatrixLayout>(std::stoi(argv[4]));
const bool do_verification = std::stoi(argv[5]);
const ck::index_t init_method = std::stoi(argv[6]);
const bool do_log = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[8]);
const float alpha = std::stof(argv[9]);
std::vector<ck::index_t> M;
std::vector<ck::index_t> N;
std::vector<ck::index_t> K;
const ck::index_t dims_arg_num = 9;
const ck::index_t dims_arg_num = 10;
collect_index_params(argv, M, dims_arg_num, 2);
collect_index_params(argv, N, dims_arg_num + 2, 2);
collect_index_params(argv, K, dims_arg_num + 4, 2);
......@@ -75,27 +77,35 @@ int profile_contraction_scale(int argc, char* argv[])
collect_index_params(argv, StridesD, dims_arg_num + 18, 4);
}
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using F64 = double;
auto profile = [&](auto a_layout, auto b_layout, auto cde_layout, auto type) {
auto profile =
[&](auto a_layout, auto b_layout, auto cde_layout, auto type, auto compute_type) {
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CDELayout = decltype(cde_layout);
using DataType = decltype(type);
using ComputeDataType = decltype(compute_type);
if(default_strides)
{
assign_default_strides(a_layout, StridesA, {M[0], M[1], K[0], K[1]});
assign_default_strides(b_layout, StridesB, {K[0], K[1], N[0], N[1]});
assign_default_strides(b_layout, StridesB, {N[0], N[1], K[0], K[1]});
assign_default_strides(cde_layout, StridesE, {M[0], M[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesD, {M[0], M[1], N[0], N[1]});
}
bool pass = ck::profiler::
profile_contraction_impl<ALayout, BLayout, CDELayout, DataType, ck::Tuple<>, Scale>(
do_verification,
bool pass = ck::profiler::profile_contraction_impl<ALayout,
BLayout,
CDELayout,
DataType,
ComputeDataType,
ck::Tuple<>,
Scale>(do_verification,
init_method,
do_log,
time_kernel,
......@@ -111,52 +121,87 @@ int profile_contraction_scale(int argc, char* argv[])
return pass;
};
if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
auto run_profile_for_datatype = [&](auto type, auto compute_type) {
if(layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
return profile(Row{}, Row{}, Row{}, F32{});
return profile(Row{}, Row{}, Row{}, type, compute_type);
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
else if(layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
return profile(Row{}, Col{}, Row{}, F32{});
return profile(Row{}, Col{}, Row{}, type, compute_type);
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
else if(layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
return profile(Col{}, Row{}, Row{}, F32{});
return profile(Col{}, Row{}, Row{}, type, compute_type);
}
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
else if(layout == ContractionMatrixLayout::KM_NK_MN_MN)
{
return profile(Col{}, Col{}, Row{}, F32{});
return profile(Col{}, Col{}, Row{}, type, compute_type);
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
return false;
};
if(data_type == ContractionDataType::F32_F32_F32_F32)
{
return profile(Row{}, Row{}, Row{}, F64{});
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F32{}, F32{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
else if(compute_data_type == ContractionComputeDataType::F16)
{
return profile(Row{}, Col{}, Row{}, F64{});
return run_profile_for_datatype(F32{}, F16{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
else if(compute_data_type == ContractionComputeDataType::BF16)
{
return profile(Col{}, Row{}, Row{}, F64{});
return run_profile_for_datatype(F32{}, BF16{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
else
{
return profile(Col{}, Col{}, Row{}, F64{});
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
else if(data_type == ContractionDataType::F64_F64_F64_F64)
{
if(compute_data_type == ContractionComputeDataType::F64)
{
return run_profile_for_datatype(F64{}, F64{});
}
else if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F64{}, F32{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
else if(data_type == ContractionDataType::F16_F16_F16_F16)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F16{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
else if(data_type == ContractionDataType::BF16_BF16_BF16_BF16)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(BF16{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
}
return 1;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_contraction_scale);
test/contraction/test_contraction.cpp
View file @ 7e8230da
......@@ -10,7 +10,10 @@
#include <gtest/gtest.h>
#include "profiler/profile_contraction_impl.hpp"
#include "profiler/profile_contraction_utils.hpp"
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using F64 = double;
......@@ -20,15 +23,11 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using Scale = ck::tensor_operation::element_wise::Scale;
struct MemoryParams
struct Dimensions
{
std::vector<ck::index_t> M;
std::vector<ck::index_t> N;
std::vector<ck::index_t> K;
std::vector<ck::index_t> StridesA;
std::vector<ck::index_t> StridesB;
std::vector<ck::index_t> StridesC;
std::vector<ck::index_t> StridesD;
};
template <typename Tuple>
......@@ -40,29 +39,33 @@ class TestContraction : public ::testing::Test
using CDLayout = std::tuple_element_t<2, Tuple>;
using DataType = std::tuple_element_t<3, Tuple>;
using DTupleDataType = std::tuple_element_t<4, Tuple>;
using CDElementOp = std::tuple_element_t<5, Tuple>;
std::vector<MemoryParams> list_of_memory_params = {{{32, 32},
{32, 32},
{32, 32},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1}},
{{16, 16},
{32, 32},
{16, 16},
{4096, 256, 16, 1},
{16, 1, 8192, 256},
{16384, 1024, 32, 1},
{16384, 1024, 32, 1}}};
std::vector<ck::index_t> init_methods = {0, 1, 2};
using ComputeDataType = std::tuple_element_t<5, Tuple>;
using CDElementOp = std::tuple_element_t<6, Tuple>;
std::vector<Dimensions> dimension_list = {{{32, 32}, {32, 32}, {32, 32}},
{{16, 16}, {32, 32}, {16, 16}}};
std::vector<ck::index_t> init_methods = {1, 2};
std::unique_ptr<CDElementOp> p_cd_element_op;
void Run()
{
for(auto& memory_params : list_of_memory_params)
for(auto& dimension_params : dimension_list)
{
std::vector<ck::index_t> StridesA;
std::vector<ck::index_t> StridesB;
std::vector<ck::index_t> StridesC;
std::vector<ck::index_t> StridesD;
const auto& M = dimension_params.M;
const auto& N = dimension_params.N;
const auto& K = dimension_params.K;
assign_default_strides(ALayout{}, StridesA, {M[0], M[1], K[0], K[1]});
assign_default_strides(BLayout{}, StridesB, {N[0], N[1], K[0], K[1]});
assign_default_strides(CDLayout{}, StridesC, {M[0], M[1], N[0], N[1]});
assign_default_strides(CDLayout{}, StridesD, {M[0], M[1], N[0], N[1]});
for(const ck::index_t init_method : init_methods)
{
bool pass =
......@@ -70,19 +73,20 @@ class TestContraction : public ::testing::Test
BLayout,
CDLayout,
DataType,
ComputeDataType,
DTupleDataType,
CDElementOp>(true /*do_verification*/,
init_method,
false /*do_logs*/,
false /*time_kernel*/,
*p_cd_element_op,
memory_params.M,
memory_params.N,
memory_params.K,
memory_params.StridesA,
memory_params.StridesB,
memory_params.StridesC,
memory_params.StridesD);
dimension_params.M,
dimension_params.N,
dimension_params.K,
StridesA,
StridesB,
StridesC,
StridesD);
EXPECT_TRUE(pass);
}
}
......@@ -99,24 +103,18 @@ class TestContractionBilinear : public TestContraction<Tuple>
{
};
#define ALL_LAYOUT_COMBINATIONS(dt, tuple_dt, compute_dt, op) \
std::tuple<Row, Row, Row, dt, tuple_dt, compute_dt, op>, \
std::tuple<Row, Col, Row, dt, tuple_dt, compute_dt, op>, \
std::tuple<Col, Row, Row, dt, tuple_dt, compute_dt, op>, \
std::tuple<Col, Col, Row, dt, tuple_dt, compute_dt, op>
using BilinearKernelTypes =
::testing::Types<std::tuple<Row, Row, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Col, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Row, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Col, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Row, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Col, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Row, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Col, Row, F64, ck::Tuple<F32>, Bilinear>>;
using ScaleKernelTypes = ::testing::Types<std::tuple<Row, Row, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Row, Col, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Col, Row, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Col, Col, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Row, Row, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Row, Col, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Col, Row, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Col, Col, Row, F64, ck::Tuple<>, Scale>>;
::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<F32>, F32, Bilinear),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<F64>, F64, Bilinear)>;
using ScaleKernelTypes = ::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<>, F32, Scale),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<>, F64, Scale)>;
TYPED_TEST_SUITE(TestContractionBilinear, BilinearKernelTypes);
TYPED_TEST_SUITE(TestContractionScale, ScaleKernelTypes);
......@@ -136,3 +134,46 @@ TYPED_TEST(TestContractionScale, scale)
this->p_cd_element_op = std::make_unique<Scale>(0.5f);
this->Run();
}
template <typename Tuple>
class TestContractionScaleMixedPrecision : public TestContraction<Tuple>
{
};
template <typename Tuple>
class TestContractionBilinearMixedPrecision : public TestContraction<Tuple>
{
};
using BilinearKernelTypesMixedPrecision =
::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<F32>, F16, Bilinear),
ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<F32>, BF16, Bilinear),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<F64>, F32, Bilinear),
ALL_LAYOUT_COMBINATIONS(F16, ck::Tuple<F16>, F32, Bilinear),
ALL_LAYOUT_COMBINATIONS(BF16, ck::Tuple<BF16>, F32, Bilinear)>;
using ScaleKernelTypesMixedPrecision =
::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<>, F16, Scale),
ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<>, BF16, Scale),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<>, F32, Scale),
ALL_LAYOUT_COMBINATIONS(F16, ck::Tuple<>, F32, Scale),
ALL_LAYOUT_COMBINATIONS(BF16, ck::Tuple<>, F32, Scale)>;
TYPED_TEST_SUITE(TestContractionBilinearMixedPrecision, BilinearKernelTypesMixedPrecision);
TYPED_TEST_SUITE(TestContractionScaleMixedPrecision, ScaleKernelTypesMixedPrecision);
TYPED_TEST(TestContractionBilinearMixedPrecision, bilinear)
{
this->p_cd_element_op = std::make_unique<Bilinear>(1.f, 1.f);
this->Run();
this->p_cd_element_op = std::make_unique<Bilinear>(-0.5f, 0.5f);
this->Run();
}
TYPED_TEST(TestContractionScaleMixedPrecision, scale)
{
this->p_cd_element_op = std::make_unique<Scale>(1.f);
this->Run();
this->p_cd_element_op = std::make_unique<Scale>(0.5f);
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