ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 04:19:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

Overhaul to Reducton and its dependants (#237)

Browse Source 
* Tiny fix in dynamic_buffer.hpp to support vectorized AtomicAdd for double type

* Update to host layer and host reduction

* Merge and remove reduction kernels

* Merge and remove reduction device interfaces and update pooling device interface

* Merge and remove useless reduction device instances

* Update to reduction profiler and reduction ctests

* Update to reduction and pooling examples and add one reduction example

* Change to reduction examples to let them testable by ctest

* Add explicit pass checking for reduction and pooling examples

* Explicit assignment of tensor shapes in example reduce_blockwise_two_call

* Use atomic_add to repace atomicAdd and add atomic_add for double type

* Add reduce ctest support for double data type

* Replace to_int_vector() by using c++ std::vector::assign()

* Keep DeviceReduceThreadWise separated from DeviceReduceBlockWise

* Merge DeviceReduceBlockWise and DeviceReduceMultiBlockAtomicAdd into DeviceReduceMultiBlock

* Add GetAtomicOperationZeroValue() support for AtomicMax

* Tiny change to reduce example README.md

* Fix some tiny issues due to branch merging

* Revoke previous change in dynamic_buffer.hpp and add atomic_add for double2_t

* Add reduce multiblock_atomic_add instances for fp64 to verify vectorized atomic_add on fp64

* Renaming

* Clean the header includings in device_reduce instances header files

[ROCm/composable_kernel commit: 63eee2d999]
This commit is contained in:
Qianfeng
2022-05-25 01:19:12 +08:00
committed by GitHub
parent 4fa2ef676a
commit 2e7ce3bb6b
94 changed files with 2429 additions and 6785 deletions
Download Patch File Download Diff File Expand all files Collapse all files

422
profiler/include/profile_reduce_impl.hpp
View File

@@ -5,74 +5,77 @@
#include "device_reduce_instance.hpp"
#include "reduction_enums.hpp"
#include "host_reduction.hpp"
#include "host_common_util.hpp"
#include "host_tensor_generator.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
template <int Rank, int NumReduceDim, int ReduceOpId, int NanOpt, int IndicesOpt>
template <int Rank, int NumReduceDim, int ReduceOpId, bool PropagateNan, bool UseIndex>
struct ReduceDescription
{
static constexpr int Rank_ = Rank;
static constexpr int NumReduceDim_ = NumReduceDim;
static constexpr int ReduceOpId_ = ReduceOpId;
static constexpr int NanOpt_ = NanOpt;
static constexpr int IndicesOpt_ = IndicesOpt;
static constexpr int PropagateNan_ = PropagateNan;
static constexpr int UseIndex_ = UseIndex;
};
using reduce_description_instances = std::tuple<ReduceDescription<4, 3, 0, 0, 0>, // for ADD
ReduceDescription<4, 4, 0, 0, 0>,
ReduceDescription<4, 1, 0, 0, 0>,
ReduceDescription<2, 1, 0, 0, 0>,
using reduce_description_instances =
std::tuple<ReduceDescription<4, 3, 0, false, false>, // for ADD
ReduceDescription<4, 4, 0, false, false>,
ReduceDescription<4, 1, 0, false, false>,
ReduceDescription<2, 1, 0, false, false>,
ReduceDescription<4, 3, 5, 0, 0>, // for AVG
ReduceDescription<4, 4, 5, 0, 0>,
ReduceDescription<4, 1, 5, 0, 0>,
ReduceDescription<2, 1, 5, 0, 0>,
ReduceDescription<4, 3, 5, false, false>, // for AVG
ReduceDescription<4, 4, 5, false, false>,
ReduceDescription<4, 1, 5, false, false>,
ReduceDescription<2, 1, 5, false, false>,
ReduceDescription<4, 3, 7, 0, 0>, // for NORM2
ReduceDescription<4, 4, 7, 0, 0>,
ReduceDescription<4, 1, 7, 0, 0>,
ReduceDescription<2, 1, 7, 0, 0>,
ReduceDescription<4, 3, 7, false, false>, // for NORM2
ReduceDescription<4, 4, 7, false, false>,
ReduceDescription<4, 1, 7, false, false>,
ReduceDescription<2, 1, 7, false, false>,
ReduceDescription<4, 3, 2, 0, 0>, // for MIN
ReduceDescription<4, 4, 2, 0, 0>,
ReduceDescription<4, 1, 2, 0, 0>,
ReduceDescription<2, 1, 2, 0, 0>,
ReduceDescription<4, 3, 3, 0, 0>, // for MAX
ReduceDescription<4, 4, 3, 0, 0>,
ReduceDescription<4, 1, 3, 0, 0>,
ReduceDescription<2, 1, 3, 0, 0>,
ReduceDescription<4, 3, 4, 0, 0>, // for AMAX
ReduceDescription<4, 4, 4, 0, 0>,
ReduceDescription<4, 1, 4, 0, 0>,
ReduceDescription<2, 1, 4, 0, 0>,
ReduceDescription<4, 3, 2, false, false>, // for MIN
ReduceDescription<4, 4, 2, false, false>,
ReduceDescription<4, 1, 2, false, false>,
ReduceDescription<2, 1, 2, false, false>,
ReduceDescription<4, 3, 3, false, false>, // for MAX
ReduceDescription<4, 4, 3, false, false>,
ReduceDescription<4, 1, 3, false, false>,
ReduceDescription<2, 1, 3, false, false>,
ReduceDescription<4, 3, 4, false, false>, // for AMAX
ReduceDescription<4, 4, 4, false, false>,
ReduceDescription<4, 1, 4, false, false>,
ReduceDescription<2, 1, 4, false, false>,
ReduceDescription<4, 3, 2, 0, 1>, // for MIN
ReduceDescription<4, 4, 2, 0, 1>,
ReduceDescription<4, 1, 2, 0, 1>,
ReduceDescription<2, 1, 2, 0, 1>,
ReduceDescription<4, 3, 3, 0, 1>, // for MAX
ReduceDescription<4, 4, 3, 0, 1>,
ReduceDescription<4, 1, 3, 0, 1>,
ReduceDescription<2, 1, 3, 0, 1>,
ReduceDescription<4, 3, 4, 0, 1>, // for AMAX
ReduceDescription<4, 4, 4, 0, 1>,
ReduceDescription<4, 1, 4, 0, 1>,
ReduceDescription<2, 1, 4, 0, 1>>;
ReduceDescription<4, 3, 2, false, true>, // for MIN
ReduceDescription<4, 4, 2, false, true>,
ReduceDescription<4, 1, 2, false, true>,
ReduceDescription<2, 1, 2, false, true>,
ReduceDescription<4, 3, 3, false, true>, // for MAX
ReduceDescription<4, 4, 3, false, true>,
ReduceDescription<4, 1, 3, false, true>,
ReduceDescription<2, 1, 3, false, true>,
ReduceDescription<4, 3, 4, false, true>, // for AMAX
ReduceDescription<4, 4, 4, false, true>,
ReduceDescription<4, 1, 4, false, true>,
ReduceDescription<2, 1, 4, false, true>>;
template <typename DescriptionType>
bool description_match(const DescriptionType& description,
int Rank,
const std::vector<int>& reduceDims,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt)
bool PropagateNan,
bool UseIndex)
{
if(description.Rank_ != Rank || description.ReduceOpId_ != static_cast<int>(ReduceOpId) ||
description.NanOpt_ != static_cast<int>(NanOpt) ||
description.IndicesOpt_ != static_cast<int>(IndicesOpt))
description.PropagateNan_ != static_cast<int>(PropagateNan) ||
description.UseIndex_ != static_cast<int>(UseIndex))
return (false);
if(DescriptionType::NumReduceDim_ != reduceDims.size())
@@ -116,46 +119,16 @@ static inline std::vector<int> get_invariant_dims(const std::vector<int>& reduce
return invariantDims;
};
template <typename T>
static void dumpBufferToFile(const char* fileName, T* data, size_t dataNumItems)
{
std::ofstream outFile(fileName, std::ios::binary);
if(outFile)
{
outFile.write(reinterpret_cast<char*>(data), dataNumItems * sizeof(T));
outFile.close();
std::cout << "Write output to file " << fileName << std::endl;
}
else
{
std::cout << "Could not open file " << fileName << " for writing" << std::endl;
}
};
// map the data type used by the GPU kernels to the corresponding type used by the host codes
template <typename InType>
struct type_mapping
{
using OutType = InType;
};
template <>
struct type_mapping<ck::half_t>
{
using OutType = half_float::half;
};
template <typename InDataType,
typename AccDataType,
typename OutDataType,
int Rank,
int NumReduceDim,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt>
void profile_reduce_impl_impl(bool do_verification,
bool PropagateNan,
bool UseIndex>
bool profile_reduce_impl_impl(bool do_verification,
int init_method,
bool do_log,
bool do_dumpout,
bool time_kernel,
const std::vector<size_t>& inLengths,
@@ -166,15 +139,13 @@ void profile_reduce_impl_impl(bool do_verification,
using namespace ck::tensor_operation::device;
using namespace ck::tensor_operation::device::device_reduce_instance;
using namespace ck::host_reduce;
using ck::host_common::dumpBufferToFile;
constexpr bool op_support_indices =
(ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
ReduceOpId == ReduceTensorOp::AMAX);
constexpr bool NeedIndices =
(op_support_indices && (IndicesOpt != ReduceTensorIndices::NO_INDICES));
constexpr bool PropagateNan = (NanOpt == NanPropagation::PROPAGATE_NAN);
constexpr bool OutputIndex = (op_support_indices && UseIndex);
constexpr bool out_support_atomic_add = std::is_same<OutDataType, float>::value;
constexpr bool op_support_atomic_add =
@@ -195,8 +166,7 @@ void profile_reduce_impl_impl(bool do_verification,
(op_support_indices && !std::is_same<AccDataType, float>::value);
// 1) The indices can only be used when the reduction operation is indexable
constexpr bool invalid_reduce_3 =
(!op_support_indices && IndicesOpt != ReduceTensorIndices::NO_INDICES);
constexpr bool invalid_reduce_3 = (!op_support_indices && UseIndex);
// 1) If InDataType is int8_t, must use int8_t as AccDataType for indexable reduction operations
// 2) If InDataType is int8_t, must use int32_t as AccDataType for non-indexable reduction
@@ -219,6 +189,8 @@ void profile_reduce_impl_impl(bool do_verification,
constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
bool pass = true;
if constexpr(!invalid_reduce)
{
Tensor<InDataType> in(inLengths);
@@ -282,7 +254,7 @@ void profile_reduce_impl_impl(bool do_verification,
if(beta != 0.0f)
out_dev.ToDevice(out.mData.data());
size_t indicesSizeInBytes = NeedIndices ? out.mDesc.GetElementSize() * sizeof(int) : 0;
size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int) : 0;
DeviceMem out_indices_dev(indicesSizeInBytes);
@@ -295,29 +267,11 @@ void profile_reduce_impl_impl(bool do_verification,
using AccElementwiseOperation_0 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
AccElementwiseOperation;
using InElementwiseOperation_1 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
InElementwiseOperation;
using AccElementwiseOperation_1 =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
AccElementwiseOperation;
using InElementwiseOperation_2 =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
InElementwiseOperation;
using AccElementwiseOperation_2 =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
AccElementwiseOperation;
using DeviceReduceInstPtr0 =
DeviceReducePtr<InElementwiseOperation_0, AccElementwiseOperation_0>;
using DeviceReduceInstPtr1 =
DeviceReducePtr<InElementwiseOperation_1, AccElementwiseOperation_1>;
using DeviceReduceInstPtr2 =
DeviceReducePtr<InElementwiseOperation_2, AccElementwiseOperation_2>;
std::vector<DeviceReduceInstPtr0> reduce0_ptrs;
std::vector<DeviceReduceInstPtr1> reduce1_ptrs;
std::vector<DeviceReduceInstPtr2> reduce2_ptrs;
add_device_reduce_instance_threadwise<InDataType,
AccDataType,
@@ -325,8 +279,8 @@ void profile_reduce_impl_impl(bool do_verification,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
PropagateNan,
UseIndex>(reduce0_ptrs);
add_device_reduce_instance_blockwise<InDataType,
AccDataType,
@@ -334,8 +288,8 @@ void profile_reduce_impl_impl(bool do_verification,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
PropagateNan,
UseIndex>(reduce0_ptrs);
if constexpr(use_atomic_add)
{
@@ -345,35 +299,11 @@ void profile_reduce_impl_impl(bool do_verification,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce0_ptrs);
PropagateNan,
UseIndex>(reduce0_ptrs);
}
else
{
add_device_reduce_instance_multiblock_partial_reduce<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce1_ptrs);
};
// used for secondary reduction
if constexpr(!use_atomic_add)
{
add_device_reduce_instance_blockwise_second_call<AccDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOpId,
NanOpt,
IndicesOpt>(reduce2_ptrs);
};
if(reduce0_ptrs.empty() && reduce1_ptrs.empty())
if(reduce0_ptrs.empty())
{
throw std::runtime_error("Wrong! No device REDUCE instance found");
};
@@ -387,23 +317,25 @@ void profile_reduce_impl_impl(bool do_verification,
Rank,
NumReduceDim,
PropagateNan,
NeedIndices>
OutputIndex>
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(
alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data());
};
const auto i_inLengths = to_int_vector(inLengths);
const auto i_inStrides = to_int_vector(inStrides);
const auto i_outLengths = to_int_vector(outLengths);
const auto i_outStrides = to_int_vector(outStrides);
std::vector<ck::index_t> i_inLengths;
std::vector<ck::index_t> i_inStrides;
std::vector<ck::index_t> i_outLengths;
std::vector<ck::index_t> i_outStrides;
i_inLengths.assign(inLengths.begin(), inLengths.end());
i_inStrides.assign(inStrides.begin(), inStrides.end());
i_outLengths.assign(outLengths.begin(), outLengths.end());
i_outStrides.assign(outStrides.begin(), outStrides.end());
for(auto& reduce_ptr : reduce0_ptrs)
{
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
DeviceMem ws_dev(wsSizeInBytes);
InElementwiseOperation_0 in_elementwise_op_0(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_0 acc_elementwise_op_0(
@@ -417,9 +349,9 @@ void profile_reduce_impl_impl(bool do_verification,
alpha,
beta,
in_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_0,
acc_elementwise_op_0);
@@ -439,8 +371,9 @@ void profile_reduce_impl_impl(bool do_verification,
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
<< std::endl;
if(time_kernel)
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< reduce_name << std::endl;
if(gb_per_sec > best_gb_per_sec)
{
@@ -450,22 +383,24 @@ void profile_reduce_impl_impl(bool do_verification,
if(do_verification)
{
out_dev.FromDevice(out.mData.data());
ck::utils::check_err(out.mData, out_ref.mData);
bool single_pass;
if(NeedIndices)
out_dev.FromDevice(out.mData.data());
single_pass = ck::utils::check_err(out.mData, out_ref.mData);
if(OutputIndex)
{
out_indices_dev.FromDevice(out_indices.mData.data());
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
;
single_pass = single_pass &&
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
};
if(do_log)
if(!single_pass)
{
LogRangeAsType<float>(std::cout << "out_host : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "out_device: ", out.mData, ",") << std::endl;
};
std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << std::endl;
}
pass = pass && single_pass;
};
if(do_dumpout)
@@ -474,7 +409,7 @@ void profile_reduce_impl_impl(bool do_verification,
dumpBufferToFile("dump_out.bin", out.mData.data(), out.mDesc.GetElementSize());
dumpBufferToFile(
"dump_out_host.bin", out_ref.mData.data(), out_ref.mDesc.GetElementSize());
if(NeedIndices)
if(OutputIndex)
{
dumpBufferToFile("dump_indices.bin",
out_indices.mData.data(),
@@ -486,158 +421,34 @@ void profile_reduce_impl_impl(bool do_verification,
};
};
for(auto& reduce_ptr : reduce1_ptrs)
{
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
DeviceMem ws_dev(wsSizeInBytes);
InElementwiseOperation_1 in_elementwise_op_1(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_1 acc_elementwise_op_1(
static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_1,
acc_elementwise_op_1);
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue;
std::string reduce_name = reduce_ptr->GetTypeString();
auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
float avg_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes =
invariant_total_length * reduce_total_length * sizeof(InDataType) +
invariant_total_length * sizeof(OutDataType);
std::vector<int> inLengths2 = reduce_ptr->GetWorkspace2dLengths(argument_ptr.get());
std::vector<int> inStrides2{inLengths2[1], 1};
for(auto& reduce2_ptr : reduce2_ptrs)
{
InElementwiseOperation_2 in_elementwise_op_2(
static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_2 acc_elementwise_op_2(
static_cast<int32_t>(reduce_total_length));
auto argument2_ptr =
reduce2_ptr->MakeArgumentPointer(inLengths2,
inStrides2,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
ws_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
in_elementwise_op_2,
acc_elementwise_op_2);
if(!reduce2_ptr->IsSupportedArgument(argument2_ptr.get()))
continue;
std::string reduce2_name = reduce2_ptr->GetTypeString();
auto invoker2_ptr = reduce2_ptr->MakeInvokerPointer();
float avg_time_2 =
invoker2_ptr->Run(argument2_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes_2 =
static_cast<size_t>(inLengths2[0]) * inLengths2[1] * sizeof(AccDataType);
float gb_per_sec = (num_bytes + num_bytes_2) / 1.E6 / (avg_time + avg_time_2);
std::cout << "Perf: " << (avg_time + avg_time_2) << " ms, " << gb_per_sec
<< " GB/s, " << reduce_name << " => " << reduce2_name << std::endl;
if(gb_per_sec > best_gb_per_sec)
{
best_avg_time = avg_time + avg_time_2;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
out_dev.FromDevice(out.mData.data());
ck::utils::check_err(out.mData, out_ref.mData);
if(NeedIndices)
{
out_indices_dev.FromDevice(out_indices.mData.data());
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
;
};
if(do_log)
{
LogRangeAsType<float>(std::cout << "out_host : ", out_ref.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "out_device: ", out.mData, ",")
<< std::endl;
}
}
if(do_dumpout)
{
dumpBufferToFile("dump_in.bin", in.mData.data(), in.mDesc.GetElementSize());
dumpBufferToFile("dump_out.bin", out.mData.data(), out.mDesc.GetElementSize());
dumpBufferToFile(
"dump_out_host.bin", out_ref.mData.data(), out_ref.mDesc.GetElementSize());
if(NeedIndices)
{
dumpBufferToFile("dump_indices.bin",
out_indices.mData.data(),
out_indices.mDesc.GetElementSize());
dumpBufferToFile("dump_indices_host.bin",
out_indices_ref.mData.data(),
out_indices_ref.mDesc.GetElementSize());
};
};
};
};
std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s"
<< std::endl;
if(time_kernel)
std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s"
<< std::endl;
}
else
{
std::cout << "The requested reduction operation is not supported, please check !!!"
<< std::endl;
};
return pass;
};
template <typename InDataType, typename AccDataType, typename OutDataType>
void profile_reduce_impl(bool do_verification,
bool profile_reduce_impl(bool do_verification,
int init_method,
bool do_log,
bool do_dumpout,
bool time_kernel,
const std::vector<size_t>& inLengths,
const std::vector<int>& reduceDims,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt,
bool PropagateNan,
bool UseIndex,
float alpha,
float beta)
{
bool matched = false;
bool pass = true;
using tuple_of_description_instances =
tensor_operation::device::device_reduce_instance::reduce_description_instances;
@@ -651,29 +462,30 @@ void profile_reduce_impl(bool do_verification,
using descType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
if(!description_match(
descType{}, inLengths.size(), reduceDims, ReduceOpId, NanOpt, IndicesOpt))
descType{}, inLengths.size(), reduceDims, ReduceOpId, PropagateNan, UseIndex))
return;
profile_reduce_impl_impl<InDataType,
AccDataType,
OutDataType,
descType::Rank_,
descType::NumReduceDim_,
static_cast<ReduceTensorOp>(descType::ReduceOpId_),
static_cast<NanPropagation>(descType::NanOpt_),
static_cast<ReduceTensorIndices>(descType::IndicesOpt_)>(
do_verification,
init_method,
do_log,
do_dumpout,
time_kernel,
inLengths,
reduceDims,
alpha,
beta);
pass = pass &&
profile_reduce_impl_impl<InDataType,
AccDataType,
OutDataType,
descType::Rank_,
descType::NumReduceDim_,
static_cast<ReduceTensorOp>(descType::ReduceOpId_),
static_cast<bool>(descType::PropagateNan_),
static_cast<bool>(descType::UseIndex_)>(do_verification,
init_method,
do_dumpout,
time_kernel,
inLengths,
reduceDims,
alpha,
beta);
matched = true;
});
return pass;
};
} // namespace profiler

218
profiler/src/profile_reduce.cpp
View File

@@ -1,27 +1,19 @@
#include <iostream>
#include <fstream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <vector>
#include <stdexcept>
#include <sstream>
#include <getopt.h>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "data_type_enum.hpp"
#include "reduction_enums.hpp"
#include "host_common_util.hpp"
#include "profile_reduce_impl.hpp"
using namespace std;
using ck::NanPropagation;
using ck::ReduceTensorIndices;
using ck::ReduceTensorOp;
static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
@@ -38,63 +30,9 @@ static struct option long_options[] = {{"inLengths", required_argument, nullptr,
{"bf16", no_argument, nullptr, '?'},
{"dumpout", required_argument, nullptr, 'o'},
{"verify", required_argument, nullptr, 'v'},
{"log", required_argument, nullptr, 'l'},
{"help", no_argument, nullptr, '?'},
{nullptr, 0, nullptr, 0}};
template <typename T>
static T getSingleValueFromString(const string& valueStr)
{
std::istringstream iss(valueStr);
T val;
iss >> val;
return (val);
};
template <typename T>
static std::vector<T> getTypeValuesFromString(const char* cstr_values)
{
std::string valuesStr(cstr_values);
std::vector<T> values;
std::size_t pos = 0;
std::size_t new_pos;
new_pos = valuesStr.find(',', pos);
while(new_pos != std::string::npos)
{
const std::string sliceStr = valuesStr.substr(pos, new_pos - pos);
T val = getSingleValueFromString<T>(sliceStr);
values.push_back(val);
pos = new_pos + 1;
new_pos = valuesStr.find(',', pos);
};
std::string sliceStr = valuesStr.substr(pos);
T val = getSingleValueFromString<T>(sliceStr);
values.push_back(val);
return (values);
}
enum struct AppDataType
{
appHalf = 0,
appFloat = 1,
appInt32 = 2,
appInt8 = 3,
appInt8x4 = 4,
appBFloat16 = 5,
appDouble = 6,
};
static void check_reduce_dims(const int rank, const std::vector<int>& reduceDims)
{
for(auto dim : reduceDims)
@@ -113,7 +51,7 @@ static void check_reduce_dims(const int rank, const std::vector<int>& reduceDims
};
};
class AppArgs
class ReduceProfilerArgs
{
private:
int option_index = 0;
@@ -130,26 +68,23 @@ class AppArgs
std::vector<float> scales;
ReduceTensorOp reduceOp = ReduceTensorOp::ADD;
AppDataType compTypeId = AppDataType::appFloat;
AppDataType outTypeId = AppDataType::appFloat;
ReduceTensorOp reduceOp = ReduceTensorOp::ADD;
ck::DataTypeEnum compTypeId = ck::DataTypeEnum::Float;
ck::DataTypeEnum outTypeId = ck::DataTypeEnum::Float;
bool compType_assigned = false;
bool outType_assigned = false;
NanPropagation nanOpt = NanPropagation::NOT_PROPAGATE_NAN;
ReduceTensorIndices indicesOpt = ReduceTensorIndices::NO_INDICES;
bool do_log = false;
bool do_verification = false;
bool do_dumpout = false;
int nanOpt = 0;
int indicesOpt = 0;
bool do_verification = false;
bool do_dumpout = false;
int init_method;
bool time_kernel;
bool need_indices = false;
AppArgs() = default;
~AppArgs() = default;
ReduceProfilerArgs() = default;
~ReduceProfilerArgs() = default;
void show_usage(const char* cmd)
{
@@ -166,8 +101,11 @@ class AppArgs
std::cout << "--outType or -W, optional enum value indicating the type of the reduced "
"output, which could be float when the input data is half"
<< std::endl;
std::cout << "--nanOpt or -N, enum value indicates the selection for NanOpt" << std::endl;
std::cout << "--indicesOpt or -I, enum value indicates the selection for IndicesOpt"
std::cout
<< "--nanOpt or -N, 1/0 value indicates the selection to use or not use Nan-Propagation"
<< std::endl;
std::cout << "--indicesOpt or -I, 1/0 value indicates the selection to use or not use "
"index in reduction"
<< std::endl;
std::cout << "--scales or -S, comma separated two float values for alpha and beta"
<< std::endl;
@@ -181,18 +119,19 @@ class AppArgs
std::cout << "--dumpout or -o, 1/0 to indicate where to save the reduction result to files "
"for further analysis"
<< std::endl;
std::cout << "--log or -l, 1/0 to indicate whether to log some information" << std::endl;
};
int processArgs(int argc, char* argv[])
{
using ck::host_common::getTypeValuesFromString;
int ch;
optind++; // to skip the "reduce" module name
while(1)
{
ch = getopt_long(argc, argv, "D:R:O:C:W:N:I:S:v:o:l:", long_options, &option_index);
ch = getopt_long(argc, argv, "D:R:O:C:W:N:I:S:v:o:", long_options, &option_index);
if(ch == -1)
break;
switch(ch)
@@ -219,27 +158,27 @@ class AppArgs
if(!optarg)
throw std::runtime_error("Invalid option format!");
compTypeId = static_cast<AppDataType>(std::atoi(optarg));
compTypeId = static_cast<ck::DataTypeEnum>(std::atoi(optarg));
compType_assigned = true;
break;
case 'W':
if(!optarg)
throw std::runtime_error("Invalid option format!");
outTypeId = static_cast<AppDataType>(std::atoi(optarg));
outTypeId = static_cast<ck::DataTypeEnum>(std::atoi(optarg));
outType_assigned = true;
break;
case 'N':
if(!optarg)
throw std::runtime_error("Invalid option format!");
nanOpt = static_cast<NanPropagation>(std::atoi(optarg));
nanOpt = std::atoi(optarg);
break;
case 'I':
if(!optarg)
throw std::runtime_error("Invalid option format!");
indicesOpt = static_cast<ReduceTensorIndices>(std::atoi(optarg));
indicesOpt = std::atoi(optarg);
break;
case 'S':
if(!optarg)
@@ -262,12 +201,6 @@ class AppArgs
do_dumpout = static_cast<bool>(std::atoi(optarg));
break;
case 'l':
if(!optarg)
throw std::runtime_error("Invalid option format!");
do_log = static_cast<bool>(std::atoi(optarg));
break;
case '?':
if(std::string(long_options[option_index].name) == "half")
use_half = true;
@@ -295,7 +228,7 @@ class AppArgs
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
init_method = std::atoi(argv[optind++]);
time_kernel = std::atoi(argv[optind]);
time_kernel = static_cast<bool>(std::atoi(argv[optind]));
if(scales.empty())
{
@@ -306,9 +239,6 @@ class AppArgs
if(reduceOp == ReduceTensorOp::MIN || reduceOp == ReduceTensorOp::MAX ||
reduceOp == ReduceTensorOp::AMAX)
{
if(indicesOpt != ReduceTensorIndices::NO_INDICES)
need_indices = true;
// for indexable operations, no need to assign compType and outType, just let them be
// same as inType
compType_assigned = false;
@@ -322,9 +252,10 @@ class AppArgs
int profile_reduce(int argc, char* argv[])
{
using namespace ck::profiler;
using ck::DataTypeEnum;
using ck::profiler::profile_reduce_impl;
AppArgs args;
ReduceProfilerArgs args;
if(args.processArgs(argc, argv) < 0)
return (-1);
@@ -339,42 +270,41 @@ int profile_reduce(int argc, char* argv[])
if(args.use_half)
{
if(!args.compType_assigned)
args.compTypeId = AppDataType::appHalf;
args.compTypeId = DataTypeEnum::Half;
if(args.outType_assigned &&
(args.outTypeId != AppDataType::appHalf && args.outTypeId != AppDataType::appFloat))
args.outTypeId = AppDataType::appFloat;
(args.outTypeId != DataTypeEnum::Half && args.outTypeId != DataTypeEnum::Float))
args.outTypeId = DataTypeEnum::Float;
if(!args.outType_assigned)
args.outTypeId = AppDataType::appHalf;
args.outTypeId = DataTypeEnum::Half;
if(args.compTypeId == AppDataType::appHalf)
if(args.compTypeId == DataTypeEnum::Half)
{
profile_reduce_impl<ck::half_t, ck::half_t, ck::half_t>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
args.scales[0],
args.scales[1]);
profile_reduce_impl<ck::half_t, ck::half_t, ck::half_t>(
args.do_verification,
args.init_method,
args.do_dumpout,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
else if(args.compTypeId == AppDataType::appFloat)
else if(args.compTypeId == DataTypeEnum::Float)
{
profile_reduce_impl<ck::half_t, float, ck::half_t>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
@@ -385,56 +315,53 @@ int profile_reduce(int argc, char* argv[])
{
profile_reduce_impl<double, double, double>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
else if(args.use_int8)
{
if(!args.compType_assigned)
args.compTypeId = AppDataType::appInt8;
args.compTypeId = DataTypeEnum::Int8;
if(args.outType_assigned &&
(args.outTypeId != AppDataType::appInt8 && args.outTypeId != AppDataType::appInt32))
args.outTypeId = AppDataType::appInt32;
(args.outTypeId != DataTypeEnum::Int8 && args.outTypeId != DataTypeEnum::Int32))
args.outTypeId = DataTypeEnum::Int32;
if(!args.outType_assigned)
args.outTypeId = AppDataType::appInt8;
args.outTypeId = DataTypeEnum::Int8;
if(args.compTypeId == AppDataType::appInt8)
if(args.compTypeId == DataTypeEnum::Int8)
{
profile_reduce_impl<int8_t, int8_t, int8_t>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
else if(args.compTypeId == AppDataType::appInt32)
else if(args.compTypeId == DataTypeEnum::Int32)
{
profile_reduce_impl<int8_t, int32_t, int8_t>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
@@ -444,54 +371,51 @@ int profile_reduce(int argc, char* argv[])
else if(args.use_bf16)
{
if(args.outType_assigned &&
(args.outTypeId != AppDataType::appBFloat16 && args.outTypeId != AppDataType::appFloat))
args.outTypeId = AppDataType::appFloat;
(args.outTypeId != DataTypeEnum::BFloat16 && args.outTypeId != DataTypeEnum::Float))
args.outTypeId = DataTypeEnum::Float;
if(!args.outType_assigned)
args.outTypeId = AppDataType::appBFloat16;
args.outTypeId = DataTypeEnum::BFloat16;
profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
else
{
if(args.compTypeId == AppDataType::appFloat)
if(args.compTypeId == DataTypeEnum::Float)
{
profile_reduce_impl<float, float, float>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}
else if(args.compTypeId == AppDataType::appDouble)
else if(args.compTypeId == DataTypeEnum::Double)
{
profile_reduce_impl<float, double, float>(args.do_verification,
args.init_method,
args.do_log,
args.do_dumpout,
args.time_kernel,
args.inLengths,
args.reduceDims,
args.reduceOp,
args.nanOpt,
args.indicesOpt,
static_cast<bool>(args.nanOpt),
static_cast<bool>(args.indicesOpt),
args.scales[0],
args.scales[1]);
}