Skip to content

GitLab

Commit bd0f0686 authored by Jing Zhang's avatar Jing Zhang
Browse files

merge develop

parents e9b1000f 63914743
Hide whitespace changes
Inline Side-by-side
Showing with 1132 additions and 655 deletions
profiler/include/profile_grouped_gemm_impl.hpp
View file @ bd0f0686
......@@ -20,7 +20,7 @@
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_grouped_gemm_instance {
namespace instance {
using DeviceGroupedGemmNoOpPtr = ck::tensor_operation::device::DeviceGroupedGemmPtr<
ck::tensor_operation::element_wise::PassThrough,
......@@ -36,7 +36,7 @@ void add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGroupedGemmNoOpPtr>&);
} // namespace device_grouped_gemm_instance
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
......@@ -171,9 +171,7 @@ void profile_grouped_gemm_impl(int do_verification,
}
// add device GEMM instances
std::vector<
ck::tensor_operation::device::device_grouped_gemm_instance::DeviceGroupedGemmNoOpPtr>
gemm_ptrs;
std::vector<ck::tensor_operation::device::instance::DeviceGroupedGemmNoOpPtr> gemm_ptrs;
if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
is_same<CDataType, half_t>::value)
......@@ -182,28 +180,28 @@ void profile_grouped_gemm_impl(int do_verification,
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
ck::tensor_operation::device::instance::
add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
ck::tensor_operation::device::instance::
add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
ck::tensor_operation::device::instance::
add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
{
ck::tensor_operation::device::device_grouped_gemm_instance::
ck::tensor_operation::device::instance::
add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
}
}
......@@ -236,6 +234,10 @@ void profile_grouped_gemm_impl(int do_verification,
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
......
profiler/include/profile_normalization_impl.hpp 0 → 100644
View file @ bd0f0686
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/conv_util.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_softmax_f16_f16_rank3_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f16_f16_rank4_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f32_f32_rank3_instances(std::vector<DeviceNormalizationPtr>&);
void add_device_softmax_f32_f32_rank4_instances(std::vector<DeviceNormalizationPtr>&);
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
enum struct NormType
{
LAYERNORM,
BATCHNORM,
SOFTMAX,
};
enum struct NormDataType
{
F32_F32, // in, out
F16_F16,
BF16_BF16,
INT8_INT8,
};
// clang-format off
template <typename NormDataType> std::string type_to_string();
template <> std::string type_to_string<float>() { return "f32"; }
template <> std::string type_to_string<half_t>() { return "f16"; }
template <> std::string type_to_string<bhalf_t>() { return "bf16"; }
template <> std::string type_to_string<int8_t>() { return "int8"; }
template <> std::string type_to_string<int32_t>() { return "int32"; }
// clang-format on
template <typename InDataType, typename AccDataType, typename OutDataType>
void profile_normalization_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> in_length,
std::vector<index_t> in_strides,
std::vector<index_t> reduce_dims,
AccDataType alpha,
AccDataType beta,
NormType norm_type)
{
Tensor<InDataType> in = in_strides.empty() ? Tensor<InDataType>(in_length)
: Tensor<InDataType>(in_length, in_strides);
Tensor<OutDataType> out(in.mDesc);
switch(init_method)
{
// case 0: break;
case 0:
in.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
out.GenerateTensorValue(GeneratorTensor_1<OutDataType>{});
break;
case 1:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break;
default:
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
}
Tensor<OutDataType> out_ref(out);
DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpace());
DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpace());
in_dev.ToDevice(in.mData.data());
out_dev.ToDevice(out.mData.data());
std::vector<index_t> i_in_lengths(in.mDesc.GetLengths().begin(), in.mDesc.GetLengths().end());
std::vector<index_t> i_in_strides(in.mDesc.GetStrides().begin(), in.mDesc.GetStrides().end());
// add device normalization instances
std::vector<tensor_operation::device::DeviceNormalizationPtr> instances;
if(norm_type == NormType::SOFTMAX)
{
if constexpr(is_same<InDataType, half_t>::value && is_same<OutDataType, half_t>::value &&
is_same<AccDataType, float>::value)
{
if(in_length.size() == 3)
tensor_operation::device::instance::add_device_softmax_f16_f16_rank3_instances(
instances);
if(in_length.size() == 4)
tensor_operation::device::instance::add_device_softmax_f16_f16_rank4_instances(
instances);
}
else if constexpr(is_same<InDataType, float>::value && is_same<OutDataType, float>::value &&
is_same<AccDataType, float>::value)
{
if(in_length.size() == 3)
tensor_operation::device::instance::add_device_softmax_f32_f32_rank3_instances(
instances);
if(in_length.size() == 4)
tensor_operation::device::instance::add_device_softmax_f32_f32_rank4_instances(
instances);
}
}
if(instances.size() <= 0)
{
throw std::runtime_error("wrong! no device normalization instance found");
}
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
for(auto& inst_ptr : instances)
{
// Is this user's responsibility to check if problem mismatches kernel instance (ie. rank 3
// problem to rank 4 kernel) other than invoking IsSupportedArgument()?
if(!(inst_ptr->GetRank() == static_cast<index_t>(i_in_lengths.size()) &&
inst_ptr->GetNumReduceDim() == static_cast<index_t>(reduce_dims.size())))
{
continue;
}
auto argument_ptr = inst_ptr->MakeArgumentPointer(i_in_lengths,
i_in_strides,
reduce_dims,
&alpha,
&beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer());
if(!inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
LogRange(std::cout << "input lengths = [", in_length, ", ")
<< "], "
<< "scaler = [" << alpha << ", " << beta << "]." << std::endl;
return;
}
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes =
in.mDesc.GetElementSize() * sizeof(InDataType) +
(beta == 0.0f ? 1 : 2) * out.mDesc.GetElementSize() * sizeof(OutDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< inst_ptr->GetTypeString() << std::endl;
if(avg_time < best_avg_time)
{
best_instance_name = inst_ptr->GetTypeString();
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
// TODO: factory method to dynamically switch between different reference normalizations
using ReferenceFactory =
tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
ReferenceFactory{}.MakeInvoker().Run({in, out_ref, alpha, beta, reduce_dims});
out_dev.FromDevice(out.mData.data());
bool pass;
if(std::is_same<InDataType, int8_t>::value)
{
pass = ck::utils::check_err(
out.mData, out_ref.mData, "Error: Incorrect results!", 0, 1);
if(do_log)
{
LogRangeAsType<int>(std::cout << "in : ", in.mData, ",") << std::endl;
LogRangeAsType<int>(std::cout << "out_ref : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<int>(std::cout << "out : ", out.mData, ",") << std::endl;
}
}
else
{
pass = ck::utils::check_err(out.mData, out_ref.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "in : ", in.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "out_ref : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "out : ", out.mData, ",") << std::endl;
}
}
if(!pass)
{
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
LogRange(std::cout << "input lengths = [", in_length, ", ")
<< "], "
<< "scaler = [" << alpha << ", " << beta << "]." << std::endl;
}
}
}
std::cout << "Best Perf for datatype = " << type_to_string<InDataType>() << "_"
<< type_to_string<OutDataType>() << ", ";
LogRange(std::cout << "length = ", i_in_lengths, ",") << ", ";
LogRange(std::cout << "stride = ", i_in_strides, ",") << ", ";
LogRange(std::cout << "reduce dims ", reduce_dims, ",") << ", ";
std::cout << "alpha = " << alpha << ", "
<< "beta = " << beta << ", " << best_avg_time << " ms, " << best_gb_per_sec
<< " GB/s, " << best_instance_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_reduce_impl.hpp
View file @ bd0f0686
......@@ -16,7 +16,7 @@
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
namespace instance {
template <int Rank, int NumReduceDim, int ReduceOpId, bool PropagateNan, bool UseIndex>
struct ReduceDescription
......@@ -91,7 +91,7 @@ bool description_match(const DescriptionType& description,
return (result);
};
} // namespace device_reduce_instance
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
......@@ -142,7 +142,7 @@ bool profile_reduce_impl_impl(bool do_verification,
float beta)
{
using namespace ck::tensor_operation::device;
using namespace ck::tensor_operation::device::device_reduce_instance;
using namespace ck::tensor_operation::device::instance;
using ck::host_common::dumpBufferToFile;
constexpr bool op_support_indices =
......@@ -464,7 +464,7 @@ bool profile_reduce_impl(bool do_verification,
bool pass = true;
using tuple_of_description_instances =
tensor_operation::device::device_reduce_instance::reduce_description_instances;
tensor_operation::device::instance::reduce_description_instances;
const auto tuple_object = tuple_of_description_instances{};
......
profiler/src/profile_batched_gemm.cpp
View file @ bd0f0686
......@@ -27,8 +27,9 @@ enum struct GemmDataType
int profile_batched_gemm(int argc, char* argv[])
{
if(argc != 15)
if(argc != 18)
{
// clang-format off
printf("arg1: tensor operation (batched_gemm: Batched GEMM)\n");
printf("arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n");
printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
......@@ -39,7 +40,8 @@ int profile_batched_gemm(int argc, char* argv[])
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n");
// clang-format on
exit(1);
}
......@@ -58,7 +60,11 @@ int profile_batched_gemm(int argc, char* argv[])
const int StrideB = std::stoi(argv[12]);
const int StrideC = std::stoi(argv[13]);
const int BatchCount = std::stoi(argv[14]);
const int BatchStrideA = std::stoi(argv[14]);
const int BatchStrideB = std::stoi(argv[15]);
const int BatchStrideC = std::stoi(argv[16]);
const int BatchCount = std::stoi(argv[17]);
using F32 = float;
using F16 = ck::half_t;
......@@ -86,6 +92,18 @@ int profile_batched_gemm(int argc, char* argv[])
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA;
const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
bool pass = ck::profiler::
profile_batched_gemm_impl<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>(
do_verification,
......@@ -95,9 +113,12 @@ int profile_batched_gemm(int argc, char* argv[])
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideC < 0) ? DefaultStrideC : StrideC,
BatchStrideA_,
BatchStrideB_,
BatchStrideC_,
StrideA_,
StrideB_,
StrideC_,
BatchCount);
return pass ? 0 : 1;
......
profiler/src/profile_convnd_bwd_weight.cpp 0 → 100644
View file @ bd0f0686
// 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_convnd_bwd_weight_impl.hpp"
namespace {
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
};
enum struct ConvInputLayout
{
NCHW, // 0
NHWC, // 1
};
enum struct ConvWeightLayout
{
KCYX, // 0
KYXC, // 1
};
enum struct ConvOutputLayout
{
NKHW, // 0
NHWK, // 1
};
ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, char* argv[], int arg_idx)
{
// (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
ck::utils::conv::ConvParams params;
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
}
} // namespace
int profile_convnd_bwd_weight(int argc, char* argv[], int num_dim_spatial)
{
const int preParams = 11;
int conv_args = 3 + num_dim_spatial * 6;
int cmdline_nargs = conv_args + preParams;
if(cmdline_nargs != argc)
{
printf("arg1: tensor operation (convnd[1|2|3]d_bwd_weight: BackwardConvolution)\n");
printf("arg2: data type (0: fp32; 1: fp16, 2: bf16)\n");
printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
printf("arg6: verification (0: no; 1: yes)\n");
printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg9: time kernel (0=n0, 1=yes)\n");
printf("arg10: splitk\n");
printf("arg11 to 25: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
return 1;
}
const auto data_type = static_cast<ConvDataType>(std::stoi(argv[2]));
const auto in_layout = static_cast<ConvInputLayout>(std::stoi(argv[3]));
const auto wei_layout = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
const auto out_layout = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const bool time_kernel = std::stoi(argv[9]);
ck::index_t split_k = std::stoi(argv[10]);
split_k = std::max(1, split_k);
ck::utils::conv::ConvParams params = parse_conv_params(num_dim_spatial, argv, preParams);
auto Run = [&](auto input_type, auto wei_type, auto out_type) {
using InDataType = decltype(input_type);
using WeiDataType = decltype(wei_type);
using OutDataType = decltype(out_type);
switch(num_dim_spatial)
{
case 1:
ck::profiler::profile_convnd_bwd_weight_impl<1,
InDataType,
WeiDataType,
OutDataType,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(
do_verification,
init_method,
do_log,
time_kernel,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.GetOutputSpatialLengths(),
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
split_k);
break;
case 2:
ck::profiler::profile_convnd_bwd_weight_impl<2,
InDataType,
WeiDataType,
OutDataType,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
time_kernel,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.GetOutputSpatialLengths(),
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
split_k);
break;
case 3:
ck::profiler::profile_convnd_bwd_weight_impl<3,
InDataType,
WeiDataType,
OutDataType,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(
do_verification,
init_method,
do_log,
time_kernel,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
params.GetOutputSpatialLengths(),
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
split_k);
break;
default: break;
}
};
if(data_type == ConvDataType::F32_F32_F32 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
Run(float{}, float{}, float{});
}
else if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
Run(ck::half_t{}, ck::half_t{}, ck::half_t{});
}
else if(data_type == ConvDataType::BF16_BF16_BF16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
Run(ck::bhalf_t{}, ck::bhalf_t{}, ck::bhalf_t{});
}
else
{
std::cout << "wrong! this Conv data_type & layout is not implemented" << std::endl;
return 1;
}
return 0;
}
profiler/src/profile_gemm_add_add_fastgelu.cpp
View file @ bd0f0686
......@@ -29,7 +29,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
if(argc != 16)
{
// clang-format off
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+GeLU)\n");
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
printf(" 1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
......@@ -39,7 +39,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
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, StrideD0, StrideD1, StrideE\n");
printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
// clang-format on
exit(1);
}
......@@ -75,9 +75,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
auto e_type,
auto a_layout,
auto b_layout,
auto d0_layout,
auto d1_layout,
auto e_layout) {
auto de_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
......@@ -87,15 +85,13 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_layout);
using D1Layout = decltype(d1_layout);
using ELayout = decltype(e_layout);
using DELayout = decltype(de_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
const int DefaultStrideD0 = ck::is_same_v<DELayout, Row> ? N : M;
const int DefaultStrideD1 = ck::is_same_v<DELayout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<DELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType,
BDataType,
......@@ -105,9 +101,7 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
EDataType,
ALayout,
BLayout,
D0Layout,
D1Layout,
ELayout>(
DELayout>(
do_verification,
init_method,
do_log,
......@@ -126,22 +120,22 @@ int profile_gemm_add_add_fastgelu(int argc, char* argv[])
if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::MK_NK_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::KM_KN_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::KM_NK_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{}, Row{});
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{});
}
else
{
......
profiler/src/profile_gemm_bias_2d.cpp deleted 100644 → 0
View file @ e9b1000f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_bias_2d_impl.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
int profile_gemm_bias_2d(int argc, char* argv[])
{
if(!(argc == 16 || argc == 17))
{
printf("arg1: tensor operation (gemm: GEMM+Bias_2d)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
printf("arg14: alpha\n");
printf("arg15: beta\n");
printf("arg16: 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 float alpha = std::stof(argv[14]);
const float beta = std::stof(argv[15]);
if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
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,
alpha,
beta);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
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) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
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) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<float,
float,
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) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_2d_impl<ck::half_t,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::ColumnMajor,
ck::tensor_layout::gemm::RowMajor>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? K : StrideA,
(StrideB < 0) ? N : StrideB,
(StrideC < 0) ? N : StrideC,
alpha,
beta);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 0;
}
profiler/src/profile_gemm_bias_relu.cpp deleted 100644 → 0
View file @ e9b1000f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_bias_relu_impl.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
int profile_gemm_bias_relu(int argc, char* argv[])
{
if(!(argc == 14 || argc == 15))
{
printf("arg1: tensor operation (gemm: GEMM+Bias+ReLU)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 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]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
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);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
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);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
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);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,
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);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 0;
}
profiler/src/profile_gemm_bias_relu_add.cpp deleted 100644 → 0
View file @ e9b1000f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "profiler/include/profile_gemm_bias_relu_add_impl.hpp"
enum struct GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum struct GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
int profile_gemm_bias_relu_add(int argc, char* argv[])
{
if(!(argc == 15 || argc == 16))
{
printf("arg1: tensor operation (gemm: GEMM+Bias+ReLU+Add)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
printf(" 1: A[m, k] * B[n, k] = C[m, n];\n");
printf(" 2: A[k, m] * B[k, n] = C[m, n];\n");
printf(" 3: A[k, m] * B[n, k] = C[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=n0, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, StrideC1\n");
printf("arg15: 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 StrideC1 = std::stoi(argv[14]);
if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
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,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
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,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
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,
(StrideC1 < 0) ? N : StrideC1);
}
else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
{
ck::profiler::profile_gemm_bias_relu_add_impl<ck::half_t,
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,
(StrideC1 < 0) ? N : StrideC1);
}
else
{
throw std::runtime_error("wrong! this data_type & layout is not implemented");
}
return 0;
}
profiler/src/profile_gemm_bilinear.cpp 0 → 100644
View file @ bd0f0686
// 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_bilinear_impl.hpp"
int profile_gemm_bilinear(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN, // 0
MK_NK_MN_MN, // 1
KM_KN_MN_MN, // 2
KM_NK_MN_MN, // 3
};
enum struct MatrixDataType
{
F32_F32_F32_F32, // 0
F16_F16_F16_F16, // 1
BF16_BF16_BF16_BF16, // 2
INT8_INT8_INT8_INT8, // 3
};
if(argc != 17)
{
// clang-format off
printf("arg1: tensor operation (gemm_bilinear: GEMM+Bilinear)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: E[m, n] = alpha * A[m, k] * B[k, n] + beta * D[m, n];\n");
printf(" 1: E[m, n] = alpha * A[m, k] * B[n, k] + beta * D[m, n];\n");
printf(" 2: E[m, n] = alpha * A[k, m] * B[k, n] + beta * D[m, n];\n");
printf(" 3: E[m, n] = alpha * A[k, m] * B[n, k] + beta * D[m, n])\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideD, StrideE\n");
printf("arg15 to 16: alhpa, beta\n");
// clang-format on
exit(1);
}
const auto data_type = static_cast<MatrixDataType>(std::stoi(argv[2]));
const auto layout = static_cast<MatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideD = std::stoi(argv[13]);
const int StrideE = std::stoi(argv[14]);
const float alpha = std::stof(argv[15]);
const float beta = std::stof(argv[16]);
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto d_type,
auto e_type,
auto a_layout,
auto b_layout,
auto de_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using DDataType = decltype(d_type);
using EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using DELayout = decltype(de_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 DefaultStrideD = ck::is_same_v<DELayout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<DELayout, Row> ? N : M;
bool pass = ck::profiler::profile_gemm_bilinear_impl<ADataType,
BDataType,
AccDataType,
DDataType,
EDataType,
ALayout,
BLayout,
DELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD < 0) ? DefaultStrideD : StrideD,
(StrideE < 0) ? DefaultStrideE : StrideE,
alpha,
beta);
return pass ? 0 : 1;
};
if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_NK_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Col{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_KN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_NK_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Col{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
}
profiler/src/profile_normalization.cpp 0 → 100644
View file @ bd0f0686
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>
#include <unordered_map>
#include "profiler/include/profile_normalization_impl.hpp"
using ck::index_t;
using ck::profiler::NormDataType;
using ck::profiler::NormType;
struct ArgParser
{
std::unordered_map<std::string, NormType> norm_dict = {{"layernorm", NormType::LAYERNORM},
{"batchnorm", NormType::BATCHNORM},
{"softmax", NormType::SOFTMAX}};
std::unordered_map<std::string, std::vector<int>> long_opts = {
{"length", {}}, {"stride", {}}, {"reduce", {}}, {"alpha", {}}, {"beta", {}}};
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()
{
std::cout << "arg1: tensor operation (layernorm/batchnorm/softmax)\n"
<< "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
<< "arg3: verification (0: no; 1: yes)\n"
<< "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
<< "arg5: print tensor value (0: no; 1: yes)\n"
<< "arg6: time kernel (0=n0, 1=yes)\n"
<< "--length: tensor extents (e.g, --length 8 4 256) \n"
<< "--stride: tensor strides (e.g, --stride 1024 256 1)\n"
<< "--reduce: to-reduce dimensions (e.g, --reduce 2)\n"
<< "--alpha: alpha scaling value\n"
<< "--beta: beta scaling value\n"
<< std::endl;
}
int profile_normalization(int argc, char* argv[])
{
if(argc <= 2)
{
print_help();
return 0;
}
ArgParser arg_parser;
// short unnamed options
const NormType norm_type = arg_parser.norm_dict[argv[1]];
const NormDataType data_type = static_cast<NormDataType>(std::stoi(argv[2]));
const bool do_verification = std::stoi(argv[3]);
const int init_method = std::stoi(argv[4]);
const bool do_log = std::stoi(argv[5]);
const bool time_kernel = std::stoi(argv[6]);
// parse the long options
arg_parser(argc, argv);
const std::vector<index_t> length = arg_parser.long_opts["length"];
const std::vector<index_t> stride = arg_parser.long_opts["stride"];
const std::vector<index_t> reduce = arg_parser.long_opts["reduce"];
const index_t alpha =
arg_parser.long_opts["alpha"].empty() ? 1 : arg_parser.long_opts["alpha"][0];
const index_t beta = arg_parser.long_opts["beta"].empty() ? 0 : arg_parser.long_opts["beta"][0];
if(data_type == NormDataType::F16_F16)
{
ck::profiler::profile_normalization_impl<ck::half_t, float, ck::half_t>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else if(data_type == NormDataType::F32_F32)
{
ck::profiler::profile_normalization_impl<float, float, float>(do_verification,
init_method,
do_log,
time_kernel,
length,
stride,
reduce,
float(alpha),
float(beta),
norm_type);
}
else
{
throw std::runtime_error("not implemented yet");
}
return 0;
}
// hijack main() for quick debugging
// int main(int argc, char* argv[])
// {
// profile_normalization(argc, argv);
// return 0;
// }
profiler/src/profiler.cpp
View file @ bd0f0686
......@@ -5,12 +5,10 @@
int profile_gemm(int, char*[]);
int profile_gemm_splitk(int, char*[]);
int profile_gemm_bias_2d(int, char*[]);
int profile_gemm_bias_relu(int, char*[]);
int profile_gemm_bias_relu_add(int, char*[]);
int profile_gemm_bias_add_reduce(int, char*[]);
int profile_gemm_bilinear(int, char*[]);
int profile_gemm_add_add_fastgelu(int, char*[]);
int profile_gemm_reduce(int, char*[]);
int profile_gemm_bias_add_reduce(int, char*[]);
int profile_batched_gemm(int, char*[]);
int profile_batched_gemm_reduce(int, char*[]);
int profile_grouped_gemm(int, char*[]);
......@@ -20,19 +18,21 @@ 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_conv_bwd_weight(int, char*[]);
int profile_normalization(int, char*[]);
int profile_reduce(int, char*[]);
int profile_convnd_bwd_weight(int, char*[], int);
static void print_helper_message()
{
// clang-format off
printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_splitk: Split-K GEMM\n"
" gemm_bias_2d: GEMM+Bias(2D)\n"
" gemm_bias_relu: GEMM+Bias+ReLU\n"
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
" gemm_bilinear: GEMM+Bilinear\n"
" gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n"
" gemm_reduce: GEMM+Reduce\n"
" gemm_bias_add_reduce: GEMM+Bias+Add+Reduce\n"
" batched_gemm: Batched GEMM\n"
" batched_gemm_reduce: Batched GEMM+Reduce\n"
" grouped_gemm: Grouped GEMM\n"
" conv_fwd: ForwardConvolution\n"
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
......@@ -62,17 +62,13 @@ int main(int argc, char* argv[])
{
return profile_gemm_splitk(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_2d") == 0)
{
return profile_gemm_bias_2d(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_relu") == 0)
else if(strcmp(argv[1], "gemm_bilinear") == 0)
{
return profile_gemm_bias_relu(argc, argv);
return profile_gemm_bilinear(argc, argv);
}
else if(strcmp(argv[1], "gemm_bias_relu_add") == 0)
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
{
return profile_gemm_bias_relu_add(argc, argv);
return profile_gemm_add_add_fastgelu(argc, argv);
}
else if(strcmp(argv[1], "gemm_reduce") == 0)
{
......@@ -118,17 +114,30 @@ int main(int argc, char* argv[])
{
return profile_convnd_bwd_data(argc, argv, 3);
}
else if(strcmp(argv[1], "reduce") == 0)
{
return profile_reduce(argc, argv);
}
else if(strcmp(argv[1], "conv2d_bwd_weight") == 0)
{
return profile_conv_bwd_weight(argc, argv);
}
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
else if(strcmp(argv[1], "convnd1d_bwd_weight") == 0)
{
return profile_gemm_add_add_fastgelu(argc, argv);
return profile_convnd_bwd_weight(argc, argv, 1);
}
else if(strcmp(argv[1], "convnd2d_bwd_weight") == 0)
{
return profile_convnd_bwd_weight(argc, argv, 2);
}
else if(strcmp(argv[1], "convnd3d_bwd_weight") == 0)
{
return profile_convnd_bwd_weight(argc, argv, 3);
}
else if(strcmp(argv[1], "reduce") == 0)
{
return profile_reduce(argc, argv);
}
else if(strcmp(argv[1], "batchnorm") == 0 || strcmp(argv[1], "layernorm") == 0 ||
strcmp(argv[1], "softmax") == 0)
{
return profile_normalization(argc, argv);
}
else
{
......
script/docker-rocm4.1.sh deleted 100755 → 0
View file @ e9b1000f
WORKSPACE=$1
echo "workspace: " $WORKSPACE
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v $WORKSPACE:/root/workspace \
rocm/tensorflow:rocm4.1-tf1.15-dev \
/bin/bash
#--network host \
script/docker-rocm4.3.1.sh deleted 100755 → 0
View file @ e9b1000f
WORKSPACE=$1
echo "workspace: " $WORKSPACE
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v $WORKSPACE:/root/workspace \
rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
#--network host \
test/CMakeLists.txt
View file @ bd0f0686
......@@ -44,6 +44,7 @@ add_subdirectory(grouped_gemm)
add_subdirectory(convnd_fwd)
add_subdirectory(reduce)
add_subdirectory(conv2d_bwd_weight)
add_subdirectory(convnd_bwd_weight)
add_subdirectory(convnd_bwd_data)
add_subdirectory(block_to_ctile_map)
add_subdirectory(softmax)
test/batched_gemm/batched_gemm_fp16.cpp
View file @ bd0f0686
......@@ -25,19 +25,19 @@ int main()
pass = pass &&
ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Row, Row, Row>(
true, 1, false, 1, M, N, K, K, N, N, BatchCount);
true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
pass = pass &&
ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Row, Col, Row>(
true, 1, false, 1, M, N, K, K, K, N, BatchCount);
true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
pass = pass &&
ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Col, Row, Row>(
true, 1, false, 1, M, N, K, M, N, N, BatchCount);
true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
pass = pass &&
ck::profiler::profile_batched_gemm_impl<ADataType, BDataType, CDataType, Col, Col, Row>(
true, 1, false, 1, M, N, K, M, K, N, BatchCount);
true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
std::cout << "test BatchedGEMM fp16: " << (pass ? "Pass" : "Fail") << std::endl;
return pass ? 0 : 1;
......
test/conv2d_bwd_data/conv2d_bwd_data.cpp
View file @ bd0f0686
......@@ -20,7 +20,7 @@ using INT8 = int8_t;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_bwd_data_instance {
namespace instance {
using DeviceConvBwdDataNoOpPtr =
DeviceConvBwdDataPtr<ck::tensor_operation::element_wise::PassThrough,
......@@ -36,7 +36,7 @@ void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
} // namespace device_conv2d_bwd_data_instance
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
......@@ -220,28 +220,28 @@ int main(int argc, char* argv[])
ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::bhalf_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, int8_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
ck::tensor_operation::device::instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
}
......
test/convnd_bwd_weight/CMakeLists.txt 0 → 100644
View file @ bd0f0686
add_test_executable(test_convnd_bwd_weight convnd_bwd_weight.cpp)
target_link_libraries(test_convnd_bwd_weight PRIVATE host_tensor device_convnd_bwd_weight_instance conv_util)
test/convnd_bwd_weight/convnd_bwd_weight.cpp 0 → 100644
View file @ bd0f0686
// 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 <vector>
#include "test/convnd_fwd/conv_util.hpp"
#include "profiler/include/profile_convnd_bwd_weight_impl.hpp"
int test_self()
{
bool pass = true;
std::vector<ck::utils::conv::ConvParams> params;
params.push_back({1, 128, 256, 256, {1}, {7}, {2}, {1}, {0}, {0}});
params.push_back({1, 128, 256, 256, {3}, {14}, {1}, {1}, {1}, {1}});
params.push_back({1, 128, 256, 256, {1}, {3}, {1}, {1}, {0}, {0}});
for(auto& param : params)
{
// f32
pass &= ck::profiler::profile_convnd_bwd_weight_impl<1,
float,
float,
float,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
// fp16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<1,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
// bf16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<1,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
}
// check 2d
params.clear();
params.push_back({2, 128, 256, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
params.push_back({2, 128, 256, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
params.push_back({2, 128, 256, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
for(auto& param : params)
{
// f32
pass &= ck::profiler::profile_convnd_bwd_weight_impl<2,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
// fp16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<2,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
// bf16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<2,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
}
// check 2d
params.clear();
params.push_back(
{3, 128, 256, 256, {1, 1, 1}, {4, 4, 4}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
params.push_back(
{3, 128, 256, 256, {3, 3, 3}, {4, 4, 8}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
params.push_back(
{3, 128, 256, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
for(auto& param : params)
{
// f32
pass &= ck::profiler::profile_convnd_bwd_weight_impl<3,
float,
float,
float,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
// fp16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<3,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
// bf16
pass &= ck::profiler::profile_convnd_bwd_weight_impl<3,
ck::bhalf_t,
ck::bhalf_t,
ck::bhalf_t,
ck::tensor_layout::convolution::NDHWC,
ck::tensor_layout::convolution::KZYXC,
ck::tensor_layout::convolution::NDHWK>(
true, // do_verification
1, // init_method
false, // do_log
true, // time_kernel
param.N_,
param.K_,
param.C_,
param.input_spatial_lengths_,
param.filter_spatial_lengths_,
param.GetOutputSpatialLengths(),
param.conv_filter_strides_,
param.conv_filter_dilations_,
param.input_left_pads_,
param.input_right_pads_,
2);
}
return pass;
}
int main()
{
// int data_type = 1;
// int init_method = 1;
bool pass = true;
pass = test_self();
if(pass)
{
std::cout << "test conv2d bwd weight : Pass" << std::endl;
return 0;
}
else
{
std::cout << "test conv2d bwd weight: Fail " << std::endl;
return -1;
}
}
test/convnd_fwd/conv_util.hpp
View file @ bd0f0686
......@@ -19,14 +19,14 @@ namespace device {
using DeviceConvFwdNoOpPtr = DeviceConvFwdPtr<element_wise::PassThrough,
element_wise::PassThrough,
element_wise::PassThrough>;
namespace device_conv2d_fwd_instance {
namespace instance {
void add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(std::vector<DeviceConvFwdNoOpPtr>&);
void add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(std::vector<DeviceConvFwdNoOpPtr>&);
} // namespace device_conv2d_fwd_instance
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
......@@ -118,7 +118,7 @@ struct ConvolutionNDFwdInstances<float, float, float>
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if(num_dim_spatial == 2)
{
ck::tensor_operation::device::device_conv2d_fwd_instance::
ck::tensor_operation::device::instance::
add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
}
return conv_ptrs;
......@@ -133,7 +133,7 @@ struct ConvolutionNDFwdInstances<ck::half_t, ck::half_t, ck::half_t>
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if(num_dim_spatial == 2)
{
ck::tensor_operation::device::device_conv2d_fwd_instance::
ck::tensor_operation::device::instance::
add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
return conv_ptrs;
......@@ -148,7 +148,7 @@ struct ConvolutionNDFwdInstances<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if(num_dim_spatial == 2)
{
ck::tensor_operation::device::device_conv2d_fwd_instance::
ck::tensor_operation::device::instance::
add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
}
return conv_ptrs;
......@@ -163,7 +163,7 @@ struct ConvolutionNDFwdInstances<int8_t, int8_t, int8_t>
std::vector<DeviceConvFwdNoOpPtr> conv_ptrs;
if(num_dim_spatial == 2)
{
ck::tensor_operation::device::device_conv2d_fwd_instance::
ck::tensor_operation::device::instance::
add_device_convnd_2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
}
return conv_ptrs;
......
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