Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "b55b52b53719cec1edf9d5351d6a74fe9a65a029"
Commit 2f463a94 authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into stream-k-initial-impl

parents ca8b5c79 ac9e01e2
Hide whitespace changes
Inline Side-by-side
Showing with 932 additions and 478 deletions
library/include/ck/library/tensor_operation_instance/gpu/pool2d_fwd.hpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto InOutRank = 4;
static constexpr auto WindowRank = 2;
static constexpr auto MaxOp = ck::ReduceTensorOp::MAX;
static constexpr auto AvgOp = ck::ReduceTensorOp::AVG;
// FP16
void add_device_pool2d_fwd_nhwc_f16_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, false>>>&);
void add_device_pool2d_fwd_nhwc_f16_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, AvgOp, false>>>&);
// FP16 - return index
void add_device_pool2d_fwd_nhwc_index_f16_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, true>>>&);
// FP32
void add_device_pool2d_fwd_nhwc_f32_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, false>>>&);
void add_device_pool2d_fwd_nhwc_f32_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, AvgOp, false>>>&);
// FP32 - return index
void add_device_pool2d_fwd_nhwc_index_f32_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, true>>>&);
template <typename InDataType,
typename OutDataType,
typename IndexDataType,
ck::ReduceTensorOp ReduceOpId,
bool OutputIndex>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
IndexDataType,
ReduceOpId,
OutputIndex>>
{
using DeviceOp = DevicePoolFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
IndexDataType,
ReduceOpId,
OutputIndex>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(is_same_v<InDataType, F16> && is_same_v<OutDataType, F16> &&
is_same_v<IndexDataType, I32>)
{
if constexpr(OutputIndex && ReduceOpId == MaxOp)
{
add_device_pool2d_fwd_nhwc_index_f16_instances(op_ptrs);
}
else
{
add_device_pool2d_fwd_nhwc_f16_instances(op_ptrs);
}
}
else if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> &&
is_same_v<IndexDataType, I32>)
{
if constexpr(OutputIndex && ReduceOpId == MaxOp)
{
add_device_pool2d_fwd_nhwc_index_f32_instances(op_ptrs);
}
else
{
add_device_pool2d_fwd_nhwc_f32_instances(op_ptrs);
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto InOutRank = 5;
static constexpr auto WindowRank = 3;
static constexpr auto MaxOp = ck::ReduceTensorOp::MAX;
static constexpr auto AvgOp = ck::ReduceTensorOp::AVG;
// FP16
void add_device_pool3d_fwd_ndhwc_f16_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, false>>>&);
void add_device_pool3d_fwd_ndhwc_f16_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, AvgOp, false>>>&);
// FP16 - return index
void add_device_pool3d_fwd_ndhwc_index_f16_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, MaxOp, true>>>&);
// FP32
void add_device_pool3d_fwd_ndhwc_f32_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, false>>>&);
void add_device_pool3d_fwd_ndhwc_f32_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, AvgOp, false>>>&);
// FP32 - return index
void add_device_pool3d_fwd_ndhwc_index_f32_instances(
std::vector<
std::unique_ptr<DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, MaxOp, true>>>&);
template <typename InDataType,
typename OutDataType,
typename IndexDataType,
ck::ReduceTensorOp ReduceOpId,
bool OutputIndex>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
IndexDataType,
ReduceOpId,
OutputIndex>>
{
using DeviceOp = DevicePoolFwd<InOutRank,
WindowRank,
InDataType,
OutDataType,
IndexDataType,
ReduceOpId,
OutputIndex>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(is_same_v<InDataType, F16> && is_same_v<OutDataType, F16> &&
is_same_v<IndexDataType, I32>)
{
if constexpr(OutputIndex && ReduceOpId == MaxOp)
{
add_device_pool3d_fwd_ndhwc_index_f16_instances(op_ptrs);
}
else
{
add_device_pool3d_fwd_ndhwc_f16_instances(op_ptrs);
}
}
else if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> &&
is_same_v<IndexDataType, I32>)
{
if constexpr(OutputIndex && ReduceOpId == MaxOp)
{
add_device_pool3d_fwd_ndhwc_index_f32_instances(op_ptrs);
}
else
{
add_device_pool3d_fwd_ndhwc_f32_instances(op_ptrs);
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
View file @ 2f463a94
...@@ -90,6 +90,7 @@ void add_device_reduce_instance_threadwise( ...@@ -90,6 +90,7 @@ void add_device_reduce_instance_threadwise(
AccElementwiseOp, AccElementwiseOp,
PropagateNan, PropagateNan,
OutputIndex, OutputIndex,
false,
false, // HaveIndexInputIfOutputIndex false, // HaveIndexInputIfOutputIndex
cfg1::BlockSize_, cfg1::BlockSize_,
cfg2::MThreadSliceSize_, cfg2::MThreadSliceSize_,
......
library/include/ck/library/utility/host_conv.hpp deleted 100644 → 0
View file @ ca8b5c79
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "host_tensor.hpp"
#include "conv_common.hpp"
template <typename TIn,
typename TWei,
typename TOut,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void host_conv_nchw_kcyx_nkhw(const Tensor<TIn>& in,
const Tensor<TWei>& wei,
Tensor<TOut>& out,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads&)
{
constexpr auto I0 = ck::Number<0>{};
constexpr auto I1 = ck::Number<1>{};
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
float v = 0;
for(int c = 0; c < wei.mDesc.GetLengths()[1]; ++c)
{
for(int y = 0; y < wei.mDesc.GetLengths()[2]; ++y)
{
int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
for(int x = 0; x < wei.mDesc.GetLengths()[3]; ++x)
{
int wi = wo * conv_strides[I1] + x * conv_dilations[I1] - in_left_pads[I1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
{
v += ck::type_convert<float>(in(n, c, hi, wi)) *
ck::type_convert<float>(wei(k, c, y, x));
}
}
}
}
out(n, k, ho, wo) = ck::type_convert<TOut>(v);
};
make_ParallelTensorFunctor(f_nchw,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2],
out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
}
template <typename TIn,
typename TWei,
typename TOut,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void host_conv3d_ndhwc_kzyxc_ndhwk(const Tensor<TIn>& in,
const Tensor<TWei>& wei,
Tensor<TOut>& out,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads&)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
const auto Di = in.mDesc.GetLengths()[1];
const auto Hi = in.mDesc.GetLengths()[2];
const auto Wi = in.mDesc.GetLengths()[3];
const auto Z = wei.mDesc.GetLengths()[1];
const auto Y = wei.mDesc.GetLengths()[2];
const auto X = wei.mDesc.GetLengths()[3];
const auto C = wei.mDesc.GetLengths()[4];
auto f_ndhwc = [&](auto n, auto do_tmp, auto ho_tmp, auto wo_tmp, auto k) {
// do__ must be converted to signed integer, otherwise zmin might be wrong in cases
// negative values.
const int do_ = static_cast<int>(do_tmp);
const int ho = static_cast<int>(ho_tmp);
const int wo = static_cast<int>(wo_tmp);
const int zmin =
std::max(0,
(in_left_pads[I0] - do_ * conv_strides[I0] + conv_dilations[I0] - 1) /
conv_dilations[I0]);
const int ymin =
std::max(0,
(in_left_pads[I1] - ho * conv_strides[I1] + conv_dilations[I1] - 1) /
conv_dilations[I1]);
const int xmin =
std::max(0,
(in_left_pads[I2] - wo * conv_strides[I2] + conv_dilations[I2] - 1) /
conv_dilations[I2]);
const int zmax =
std::min(Z, (in_left_pads[I0] - do_ * conv_strides[I0] + Di) / conv_dilations[I0]);
const int ymax =
std::min(Y, (in_left_pads[I1] - ho * conv_strides[I1] + Hi) / conv_dilations[I1]);
const int xmax =
std::min(X, (in_left_pads[I2] - wo * conv_strides[I2] + Wi) / conv_dilations[I2]);
const int di_min = do_ * conv_strides[I0] + zmin * conv_dilations[I0] - in_left_pads[I0];
const int hi_min = ho * conv_strides[I1] + ymin * conv_dilations[I1] - in_left_pads[I1];
const int wi_min = wo * conv_strides[I2] + xmin * conv_dilations[I2] - in_left_pads[I2];
double v = 0;
const TIn* in_n = in.mData.data() + n * Di * Hi * Wi * C;
const TWei* wei_k = wei.mData.data() + k * Z * Y * X * C;
int di = di_min;
for(int z = zmin; z < zmax; ++z, di += conv_dilations[I0])
{
const TIn* in_n_di = in_n + di * Hi * Wi * C;
const TWei* wei_k_z = wei_k + z * Y * X * C;
int hi = hi_min;
for(int y = ymin; y < ymax; ++y, hi += conv_dilations[I1])
{
const TIn* in_n_di_hi = in_n_di + hi * Wi * C;
const TWei* wei_k_z_y = wei_k_z + y * X * C;
int wi = wi_min;
for(int x = xmin; x < xmax; ++x, wi += conv_dilations[I2])
{
const TIn* in_n_di_hi_wi = in_n_di_hi + wi * C;
const TWei* wei_k_z_y_x = wei_k_z_y + x * C;
for(int c = 0; c < C; ++c)
{
v += static_cast<const double>(in_n_di_hi_wi[c]) *
static_cast<const double>(wei_k_z_y_x[c]);
}
}
}
}
out(n, do_, ho, wo, k) = v;
};
make_ParallelTensorFunctor(f_ndhwc,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2],
out.mDesc.GetLengths()[3],
out.mDesc.GetLengths()[4])(std::thread::hardware_concurrency() - 4);
}
library/include/ck/library/utility/host_tensor.hpp
View file @ 2f463a94
...@@ -411,6 +411,12 @@ struct Tensor ...@@ -411,6 +411,12 @@ struct Tensor
} }
} }
template <typename... Is>
std::size_t GetOffsetFromMultiIndex(Is... is) const
{
return mDesc.GetOffsetFromMultiIndex(is...);
}
template <typename... Is> template <typename... Is>
T& operator()(Is... is) T& operator()(Is... is)
{ {
......
library/include/ck/library/utility/op_instance_engine.hpp deleted 100644 → 0
View file @ ca8b5c79
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <cstdlib>
#include <iostream>
#include <limits>
#include <memory>
#include <stdexcept>
#include <tuple>
#include <utility>
#include <vector>
#include "ck/utility/functional2.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
namespace ck {
namespace utils {
struct ProfileBestConfig
{
std::string best_op_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_tflops = std::numeric_limits<float>::max();
float best_gb_per_sec = std::numeric_limits<float>::max();
};
/**
* @brief This class describes an operation instance(s).
*
* Op instance defines a particular specializations of operator
* template. Thanks to this specific input/output data types, data
* layouts and modifying elementwise operations it is able to create
* it's input/output tensors, provide pointers to instances which
* can execute it and all operation specific parameters.
*/
template <typename OutDataType, typename... InArgTypes>
class OpInstance
{
public:
template <typename T>
using TensorPtr = std::unique_ptr<Tensor<T>>;
using InTensorsTuple = std::tuple<TensorPtr<InArgTypes>...>;
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
using DeviceBuffers = std::vector<DeviceMemPtr>;
OpInstance() = default;
OpInstance(const OpInstance&) = default;
OpInstance& operator=(const OpInstance&) = default;
virtual ~OpInstance(){};
virtual InTensorsTuple GetInputTensors() const = 0;
virtual TensorPtr<OutDataType> GetOutputTensor() const = 0;
virtual std::unique_ptr<tensor_operation::device::BaseInvoker>
MakeInvokerPointer(tensor_operation::device::BaseOperator*) const = 0;
virtual std::unique_ptr<tensor_operation::device::BaseArgument>
MakeArgumentPointer(tensor_operation::device::BaseOperator*,
const DeviceBuffers&,
const DeviceMemPtr&) const = 0;
virtual std::size_t GetFlops() const = 0;
virtual std::size_t GetBtype() const = 0;
};
/**
* @brief A generic operation instance run engine.
*/
template <typename OutDataType, typename... InArgTypes>
class OpInstanceRunEngine
{
public:
using OpInstanceT = OpInstance<InArgTypes..., OutDataType>;
template <typename T>
using TensorPtr = std::unique_ptr<Tensor<T>>;
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
using InTensorsTuple = std::tuple<TensorPtr<InArgTypes>...>;
using DeviceBuffers = std::vector<DeviceMemPtr>;
using InArgsTypesTuple = std::tuple<InArgTypes...>;
OpInstanceRunEngine() = delete;
template <typename ReferenceOp = std::function<void()>>
OpInstanceRunEngine(const OpInstanceT& op_instance,
const ReferenceOp& reference_op = ReferenceOp{},
bool do_verification = true)
: op_instance_{op_instance}
{
in_tensors_ = op_instance_.GetInputTensors();
out_tensor_ = op_instance_.GetOutputTensor();
if constexpr(std::is_invocable_v<ReferenceOp,
const Tensor<InArgTypes>&...,
Tensor<OutDataType>&>)
{
if(do_verification)
{
ref_output_ = op_instance_.GetOutputTensor();
CallRefOpUnpackArgs(reference_op, std::make_index_sequence<kNInArgs_>{});
}
}
AllocateDeviceInputTensors(std::make_index_sequence<kNInArgs_>{});
out_device_buffer_ = std::make_unique<DeviceMem>(sizeof(OutDataType) *
out_tensor_->mDesc.GetElementSpaceSize());
out_device_buffer_->SetZero();
}
virtual ~OpInstanceRunEngine(){};
template <typename OpInstancePtr>
bool Test(const std::vector<OpInstancePtr>& op_ptrs)
{
bool res{true};
for(auto& op_ptr : op_ptrs)
{
auto invoker = op_instance_.MakeInvokerPointer(op_ptr.get());
auto argument = op_instance_.MakeArgumentPointer(
op_ptr.get(), in_device_buffers_, out_device_buffer_);
if(op_ptr->IsSupportedArgument(argument.get()))
{
std::cout << "Testing instance: " << op_ptr->GetTypeString() << std::endl;
invoker->Run(argument.get());
out_device_buffer_->FromDevice(out_tensor_->mData.data());
if(!ref_output_)
{
throw std::runtime_error(
"OpInstanceRunEngine::Test: Reference value not availabe."
" You have to provide reference function.");
}
// TODO: enable flexible use of custom check_error functions
bool inst_res = CheckErr(out_tensor_->mData, ref_output_->mData);
std::cout << (inst_res ? "SUCCESS" : "FAILURE") << std::endl;
res = res && inst_res;
out_device_buffer_->SetZero();
}
else
{
std::cout << "Given conv problem is not supported by instance: \n\t>>>>"
<< op_ptr->GetTypeString() << std::endl;
}
}
return res;
}
template <typename OpInstancePtr>
ProfileBestConfig Profile(const std::vector<OpInstancePtr>& op_ptrs,
bool time_kernel = false,
bool do_verification = false,
bool do_log = false)
{
ProfileBestConfig best_config;
for(auto& op_ptr : op_ptrs)
{
auto invoker = op_instance_.MakeInvokerPointer(op_ptr.get());
auto argument = op_instance_.MakeArgumentPointer(
op_ptr.get(), in_device_buffers_, out_device_buffer_);
if(op_ptr->IsSupportedArgument(argument.get()))
{
std::string op_name = op_ptr->GetTypeString();
float avg_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flops = op_instance_.GetFlops();
std::size_t num_btype = op_instance_.GetBtype();
float tflops = static_cast<float>(flops) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << op_name << std::endl;
if(avg_time < best_config.best_avg_time)
{
best_config.best_op_name = op_name;
best_config.best_tflops = tflops;
best_config.best_gb_per_sec = gb_per_sec;
best_config.best_avg_time = avg_time;
}
if(do_verification)
{
out_device_buffer_->FromDevice(out_tensor_->mData.data());
if(!ref_output_)
{
throw std::runtime_error(
"OpInstanceRunEngine::Profile: Reference value not availabe."
" You have to provide reference function.");
}
// TODO: enable flexible use of custom check_error functions
CheckErr(out_tensor_->mData, ref_output_->mData);
if(do_log) {}
}
out_device_buffer_->SetZero();
}
}
return best_config;
}
void SetAtol(double a) { atol_ = a; }
void SetRtol(double r) { rtol_ = r; }
private:
template <typename F, std::size_t... Is>
void CallRefOpUnpackArgs(const F& f, std::index_sequence<Is...>) const
{
f(*std::get<Is>(in_tensors_)..., *ref_output_);
}
template <std::size_t... Is>
void AllocateDeviceInputTensors(std::index_sequence<Is...>)
{
(AllocateDeviceInputTensorsImpl<Is>(), ...);
}
template <std::size_t Index>
void AllocateDeviceInputTensorsImpl()
{
const auto& ts = std::get<Index>(in_tensors_);
in_device_buffers_
.emplace_back(
std::make_unique<DeviceMem>(sizeof(std::tuple_element_t<Index, InArgsTypesTuple>) *
ts->mDesc.GetElementSpaceSize()))
->ToDevice(ts->mData.data());
}
static constexpr std::size_t kNInArgs_ = std::tuple_size_v<InTensorsTuple>;
const OpInstanceT& op_instance_;
double rtol_{1e-5};
double atol_{1e-8};
InTensorsTuple in_tensors_;
TensorPtr<OutDataType> out_tensor_;
TensorPtr<OutDataType> ref_output_;
DeviceBuffers in_device_buffers_;
DeviceMemPtr out_device_buffer_;
template <typename T>
bool CheckErr(const std::vector<T>& dev_out, const std::vector<T>& ref_out) const
{
return ck::utils::check_err(dev_out, ref_out, "Error: incorrect results!", rtol_, atol_);
}
};
} // namespace utils
} // namespace ck
library/src/tensor_operation_instance/gpu/pool_fwd/CMakeLists.txt 0 → 100644
View file @ 2f463a94
add_instance_library(device_pool_fwd_instance
device_avg_pool2d_fwd_nhwc_f16_instance.cpp
device_avg_pool2d_fwd_nhwc_f32_instance.cpp
device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
device_max_pool2d_fwd_nhwc_f16_instance.cpp
device_max_pool2d_fwd_nhwc_f32_instance.cpp
device_max_pool3d_fwd_ndhwc_f16_instance.cpp
device_max_pool3d_fwd_ndhwc_f32_instance.cpp
)
library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool2d_fwd_nhwc_f16_instance.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "pool_fwd_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
void add_device_pool2d_fwd_nhwc_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F16, F16, I32, ReduceOpId, false>>>& instances)
{
add_device_operation_instances(
instances, device_pool2d_fwd_nhwc_instances<F16, F16, I32, F32, ReduceOpId, false>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool2d_fwd_nhwc_f32_instance.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "pool_fwd_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
void add_device_pool2d_fwd_nhwc_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, ReduceOpId, false>>>& instances)
{
add_device_operation_instances(
instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "pool_fwd_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
void add_device_pool3d_fwd_ndhwc_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, ReduceOpId, false>>>& instances)
{
add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F32, ReduceOpId, false>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "pool_fwd_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
void add_device_pool3d_fwd_ndhwc_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, false>>>& instances)
{
add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool2d_fwd_nhwc_f16_instance.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "pool_fwd_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
void add_device_pool2d_fwd_nhwc_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F16, F16, I32, ReduceOpId, false>>>& instances)
{
add_device_operation_instances(
instances, device_pool2d_fwd_nhwc_instances<F16, F16, I32, F16, ReduceOpId, false>{});
}
void add_device_pool2d_fwd_nhwc_index_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F16, F16, I32, ReduceOpId, true>>>& instances)
{
add_device_operation_instances(
instances, device_pool2d_fwd_nhwc_instances<F16, F16, I32, F16, ReduceOpId, true>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool2d_fwd_nhwc_f32_instance.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "pool_fwd_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
void add_device_pool2d_fwd_nhwc_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, ReduceOpId, false>>>& instances)
{
add_device_operation_instances(
instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
}
void add_device_pool2d_fwd_nhwc_index_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, ReduceOpId, true>>>& instances)
{
add_device_operation_instances(
instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, true>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "pool_fwd_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
void add_device_pool3d_fwd_ndhwc_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, ReduceOpId, false>>>& instances)
{
add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F16, ReduceOpId, false>{});
}
void add_device_pool3d_fwd_ndhwc_index_f16_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, ReduceOpId, true>>>& instances)
{
add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F16, ReduceOpId, true>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "pool_fwd_instance_common.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
void add_device_pool3d_fwd_ndhwc_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, false>>>& instances)
{
add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
}
void add_device_pool3d_fwd_ndhwc_index_f32_instances(
std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, true>>>& instances)
{
add_device_operation_instances(
instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, true>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/pool_fwd/pool_fwd_instance_common.hpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using I32 = int32_t;
using F16 = ck::half_t;
using F32 = float;
template <typename InDataType,
typename OutDataType,
typename IndexDataType,
typename ComputeDataType,
ReduceTensorOp ReduceOpId,
bool OutputIndex>
using device_pool2d_fwd_nhwc_instances =
// clang-format off
std::tuple <
DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
// clang-format on
>;
template <typename InDataType,
typename OutDataType,
typename IndexDataType,
typename ComputeDataType,
ReduceTensorOp ReduceOpId,
bool OutputIndex>
using device_pool3d_fwd_ndhwc_instances =
// clang-format off
std::tuple <
DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/README.md
View file @ 2f463a94
...@@ -46,3 +46,33 @@ out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256} ...@@ -46,3 +46,33 @@ out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
.... ....
Best Perf: 1.42509 ms, 102.988 TFlops, 234.086 GB/s 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];
# 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
```
Result (MI100)
```bash
a_m_k: dim 4, lengths {128, 128, 128, 128}, strides {2097152, 16384, 128, 1}
b_k_n: dim 4, lengths {128, 128, 128, 128}, strides {128, 1, 2097152, 16384}
d_m_n: dim 4, lengths {128, 128, 128, 128}, strides {2097152, 16384, 128, 1}
e_m_n: dim 4, lengths {128, 128, 128, 128}, strides {2097152, 16384, 128, 1}
....
Best Perf: 211.405 ms, 41.6077 TFlops, 15.2372 GB/s
```
profiler/include/profiler/data_type_enum_helper.hpp deleted 100644 → 0
View file @ ca8b5c79
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma
#include "ck/utility/data_type.hpp"
#include "profiler/data_type_enum.hpp"
namespace ck {
template <DataTypeEnum DataTypeEnum>
struct get_datatype_from_enum;
template <>
struct get_datatype_from_enum<DataTypeEnum::Int8>
{
using type = int8_t;
};
template <>
struct get_datatype_from_enum<DataTypeEnum::Int32>
{
using type = int32_t;
};
template <>
struct get_datatype_from_enum<DataTypeEnum::Half>
{
using type = half_t;
};
template <>
struct get_datatype_from_enum<DataTypeEnum::Float>
{
using type = float;
};
template <>
struct get_datatype_from_enum<DataTypeEnum::Double>
{
using type = double;
};
template <typename T>
struct get_datatype_enum_from_type;
template <>
struct get_datatype_enum_from_type<int8_t>
{
static constexpr DataTypeEnum value = DataTypeEnum::Int8;
};
template <>
struct get_datatype_enum_from_type<int32_t>
{
static constexpr DataTypeEnum value = DataTypeEnum::Int32;
};
template <>
struct get_datatype_enum_from_type<half_t>
{
static constexpr DataTypeEnum value = DataTypeEnum::Half;
};
template <>
struct get_datatype_enum_from_type<float>
{
static constexpr DataTypeEnum value = DataTypeEnum::Float;
};
template <>
struct get_datatype_enum_from_type<double>
{
static constexpr DataTypeEnum value = DataTypeEnum::Double;
};
} // namespace ck
profiler/include/profiler/profile_contraction_impl.hpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include <iostream>
#include <typeinfo>
#include <limits>
#include <vector>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
#include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
#include "ck/host_utility/io.hpp"
namespace ck {
namespace profiler {
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using Scale = ck::tensor_operation::element_wise::Scale;
template <typename ALayout,
typename BLayout,
typename CDELayout,
typename DataType,
typename DTupleDataType,
typename CDElementOp>
int profile_contraction_impl(ck::index_t do_verification,
ck::index_t init_method,
bool do_log,
bool time_kernel,
CDElementOp cde_element_op,
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)
{
bool pass = true;
auto f_host_tensor_descriptor = [](const std::vector<ck::index_t>& dims01,
const std::vector<ck::index_t>& dims23,
const std::vector<ck::index_t>& strides) {
std::vector<std::size_t> dims_szt(dims01.begin(), dims01.end());
dims_szt.insert(dims_szt.end(), dims23.begin(), dims23.end());
std::vector<std::size_t> strides_szt(strides.begin(), strides.end());
return HostTensorDescriptor(dims_szt, strides);
};
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> 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 << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
b_k_n.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});
d_m_n.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5});
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
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 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());
e_device_buf.SetZero();
d_device_buf.ToDevice(d_m_n.mData.data());
const std::vector<index_t> a_ms_ks_lengths = {M[0], M[1], K[0], K[1]};
const std::vector<index_t> b_ns_ks_lengths = {N[0], N[1], K[0], K[1]};
const std::vector<index_t> e_ms_ns_lengths = {M[0], M[1], N[0], N[1]};
const std::vector<index_t> d_m_n_lengths = {M[0], M[1], N[0], N[1]};
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
constexpr ck::index_t NumDim = 2;
using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<NumDim,
NumDim,
NumDim,
DataType,
DataType,
DTupleDataType,
DataType,
AElementOp,
BElementOp,
CDElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// Run reference op
if(do_verification)
{
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDim,
NumDim,
NumDim,
DataType,
DataType,
DataType,
DataType,
AElementOp,
BElementOp>;
auto ref_op = ReferenceGemmInstance{};
auto ref_invoker = ref_op.MakeInvoker();
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_invoker.Run(ref_argument);
for(size_t m0 = 0; m0 < e_m_n_host_result.mDesc.GetLengths()[0]; ++m0)
{
for(size_t m1 = 0; m1 < e_m_n_host_result.mDesc.GetLengths()[1]; ++m1)
{
for(size_t n0 = 0; n0 < e_m_n_host_result.mDesc.GetLengths()[2]; ++n0)
{
for(size_t n1 = 0; n1 < e_m_n_host_result.mDesc.GetLengths()[3]; ++n1)
{
if constexpr(is_same<CDElementOp, Bilinear>::value)
{
cde_element_op(e_m_n_host_result(m0, m1, n0, n1),
c_m_n_host_result(m0, m1, n0, n1),
d_m_n(m0, m1, n0, n1));
}
else if constexpr(is_same<CDElementOp, Scale>::value)
{
cde_element_op(e_m_n_host_result(m0, m1, n0, n1),
c_m_n_host_result(m0, m1, n0, n1));
}
else
{
static_assert("Unsupported CDElementOp in contraction profiler.");
}
}
}
}
}
}
std::string best_op_name;
float best_avg_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device op instances
for(auto& op_ptr : op_ptrs)
{
std::unique_ptr<tensor_operation::device::BaseArgument> argument_ptr;
if constexpr(is_same<CDElementOp, Bilinear>::value)
{
argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<DataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(b_device_buf.GetDeviceBuffer()),
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
static_cast<DataType*>(e_device_buf.GetDeviceBuffer()),
a_ms_ks_lengths,
StridesA,
b_ns_ks_lengths,
StridesB,
std::array<std::vector<ck::index_t>, 1>{d_m_n_lengths},
std::array<std::vector<ck::index_t>, 1>{StridesD},
e_ms_ns_lengths,
StridesE,
a_element_op,
b_element_op,
cde_element_op);
}
else if constexpr(is_same<CDElementOp, Scale>::value)
{
argument_ptr =
op_ptr->MakeArgumentPointer(static_cast<DataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<DataType*>(b_device_buf.GetDeviceBuffer()),
std::array<const void*, 0>{},
static_cast<DataType*>(e_device_buf.GetDeviceBuffer()),
a_ms_ks_lengths,
StridesA,
b_ns_ks_lengths,
StridesB,
std::array<std::vector<ck::index_t>, 0>{},
std::array<std::vector<ck::index_t>, 0>{},
e_ms_ns_lengths,
StridesE,
a_element_op,
b_element_op,
cde_element_op);
}
else
{
static_assert("Unsupported CDElementOp in contraction profiler.");
}
auto invoker_ptr = op_ptr->MakeInvokerPointer();
auto nelems_m = M[0] * M[1];
auto nelems_n = N[0] * N[1];
auto nelems_k = K[0] * K[1];
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init C to zero before profiling next kernel
e_device_buf.SetZero();
std::string op_name = op_ptr->GetTypeString();
float avg_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * nelems_m * nelems_n * nelems_k;
std::size_t num_btype = sizeof(DataType) * nelems_m * nelems_k +
sizeof(DataType) * nelems_k * nelems_n +
sizeof(DataType) * nelems_m * nelems_n;
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(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();
pass = pass & ck::utils::check_err(e_m_n_device_result,
e_m_n_host_result,
"Error: incorrect results!",
threshold,
threshold);
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 << "c_host : ", e_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", e_m_n_device_result.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
}
}
if constexpr(is_same<DataType, float>::value)
{
std::cout << "Best Perf for datatype = f32";
}
else if constexpr(is_same<DataType, double>::value)
{
std::cout << "Best Perf for datatype = f64";
}
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " ALayout = RowMajor";
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " ALayout = ColumnMajor";
}
if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " BLayout = RowMajor";
}
else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " BLayout = ColumnMajor";
}
if constexpr(is_same<CDELayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " CDELayout = RowMajor";
}
else if constexpr(is_same<CDELayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " CDELayout = ColumnMajor";
}
std::cout << " M = " << M << " N = " << N << " K = " << K << " StridesA = " << StridesA
<< " StridesB = " << StridesB << " StridesE = " << StridesE << " : " << best_avg_time
<< " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
<< best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_contraction_utils.hpp 0 → 100644
View file @ 2f463a94
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include "ck/ck.hpp"
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using Scale = ck::tensor_operation::element_wise::Scale;
enum struct ContractionMatrixLayout
{
MK_KN_MN_MN, // 0
MK_NK_MN_MN, // 1
KM_KN_MN_MN, // 2
KM_NK_MN_MN, // 3
};
enum struct ContractionDataType
{
F32_F32_F32_F32, // 0
F64_F64_F64_F64, // 1
};
inline void collect_index_params(char* argv[],
std::vector<ck::index_t>& params,
const ck::index_t from,
const ck::index_t num)
{
for(ck::index_t p = from; p < from + num; p++)
params.push_back(std::stoi(argv[p]));
}
// Defualt strides for row-major: {Dim1 * Dim2 * Dim3, Dim2 * Dim3, Dim3, 1}
// Defualt strides for column-major: {Dim1, 1, Dim0 * Dim1 * Dim3, Dim0 * Dim1}
inline void
assign_default_strides(Row, std::vector<ck::index_t>& strides, std::vector<ck::index_t> dims)
{
strides = {dims[1] * dims[2] * dims[3], dims[2] * dims[3], dims[3], 1};
}
inline void
assign_default_strides(Col, std::vector<ck::index_t>& strides, std::vector<ck::index_t> dims)
{
strides = {dims[1], 1, dims[0] * dims[1] * dims[3], dims[0] * dims[1]};
}
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