ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-23 22:34:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

Unify the naming of the math functions used by the host and kernel (#262)

Browse Source 
* Use the unified naming for math functions on host and HIP kernel

* Corresponding change/simplification in reduction host/profiler/examples due to unified math functions renaming

* Renaming GetReductionZeroVal() to GetIdentityValue()

* Tiny renaming in profile_reduce_impl.hpp

* More renaming in profile_reduce_impl.hpp

* Replace zeroVal by identiyVal

* Remove ck_ prefix in the naming of ck::math provided functions

[ROCm/composable_kernel commit: 86185bd7ce]
This commit is contained in:
Qianfeng
2022-06-02 10:49:53 +08:00
committed by GitHub
parent 8d5fe58cd5
commit 3c29c7cac5
22 changed files with 198 additions and 417 deletions
Download Patch File Download Diff File Expand all files Collapse all files

257
library/include/ck/library/host_tensor/host_reduce_util.hpp
View File

@@ -1,257 +0,0 @@
/*******************************************************************************
*
* MIT License
*
* Copyright (c) 2020 Advanced Micro Devices, Inc.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
*******************************************************************************/
#ifndef GUARD_HOST_REDUCE_UTIL_HPP
#define GUARD_HOST_REDUCE_UTIL_HPP
#include <limits>
#include <cmath>
#include <functional>
#include "reduction_enums.hpp"
#include "data_type.hpp"
#include "math_v2.hpp"
namespace ck {
namespace host_reduce {
using ck::NanPropagation;
using ck::ReduceTensorOp;
template <typename AccDataType, ReduceTensorOp ReduceOpId>
__host__ static inline std::function<void(AccDataType&)> PreUnaryOpFn(int)
{
using ck::math::abs;
if constexpr(ReduceOpId == ReduceTensorOp::NORM1)
{
return ([&](AccDataType& a_) { a_ = abs(a_); });
}
else if constexpr(ReduceOpId == ReduceTensorOp::NORM2)
{
return ([&](AccDataType& a_) { a_ = a_ * a_; });
}
else if constexpr(ReduceOpId == ReduceTensorOp::AMAX)
{
return ([&](AccDataType& a_) { a_ = abs(a_); });
}
else
{
// ReduceTensorOp::AVG:
// ReduceTensorOp::ADD:
// ReduceTensorOp::MUL:
// ReduceTensorOp::MIN:
// ReduceTensorOp::MAX:
return ([&](AccDataType&) {});
};
};
template <typename AccDataType, ReduceTensorOp ReduceOpId>
__host__ static inline std::function<void(AccDataType&)> PosUnaryOpFn(int32_t divider)
{
using std::sqrt;
if constexpr(ReduceOpId == ReduceTensorOp::NORM2)
{
return ([&](AccDataType& a_) { a_ = sqrt(a_); });
}
else if constexpr(ReduceOpId == ReduceTensorOp::AVG)
{
return ([&, divider](AccDataType& a_) {
a_ = a_ / static_cast<AccDataType>(static_cast<float>(divider));
});
}
else
{
// ReduceTensorOp::ADD:
// ReduceTensorOp::NORM1:
// ReduceTensorOp::MUL:
// ReduceTensorOp::MIN:
// ReduceTensorOp::MAX:
// ReduceTensorOp::AMAX:
return ([&](AccDataType&) {});
}
};
template <typename AccDataType, ReduceTensorOp ReduceOpId>
__host__ static inline std::function<void(AccDataType&, AccDataType)> ReduceOpFn()
{
if constexpr(ReduceOpId == ReduceTensorOp::ADD || ReduceOpId == ReduceTensorOp::AVG ||
ReduceOpId == ReduceTensorOp::NORM1 || ReduceOpId == ReduceTensorOp::NORM2)
{
return ([&](AccDataType& a_, AccDataType b_) { a_ = a_ + b_; });
}
else if constexpr(ReduceOpId == ReduceTensorOp::MUL)
{
return ([&](AccDataType& a_, AccDataType b_) { a_ = a_ * b_; });
}
else if constexpr(ReduceOpId == ReduceTensorOp::MIN)
{
return ([&](AccDataType& a_, AccDataType b_) {
if(a_ > b_)
a_ = b_;
});
}
else if constexpr(ReduceOpId == ReduceTensorOp::MAX || ReduceOpId == ReduceTensorOp::AMAX)
{
return ([&](AccDataType& a_, AccDataType b_) {
if(a_ < b_)
a_ = b_;
});
}
};
template <typename AccDataType, ReduceTensorOp ReduceOpId>
__host__ static inline std::function<void(AccDataType&, AccDataType, bool& changed)> ReduceOpFn2()
{
if constexpr(ReduceOpId == ReduceTensorOp::MIN)
{
return ([&](AccDataType& a_, AccDataType b_, bool& changed) {
if(a_ > b_)
{
a_ = b_;
changed = true;
}
else
changed = false;
});
}
else if constexpr(ReduceOpId == ReduceTensorOp::MAX || ReduceOpId == ReduceTensorOp::AMAX)
{
return ([&](AccDataType& a_, AccDataType b_, bool& changed) {
if(a_ < b_)
{
a_ = b_;
changed = true;
}
else
changed = false;
});
}
else
{
// ReduceTensorOp::ADD:
// ReduceTensorOp::MUL:
// ReduceTensorOp::AVG:
// ReduceTensorOp::NORM1:
// ReduceTensorOp::NORM2:
return (std::function<void(AccDataType&, AccDataType, bool&)>{});
};
};
template <typename AccDataType, ReduceTensorOp ReduceOpId>
__host__ static inline AccDataType ReduceOpZeroVal()
{
if constexpr(ReduceOpId == ReduceTensorOp::MUL)
{
return (static_cast<AccDataType>(1.0f));
}
else if constexpr(ReduceOpId == ReduceTensorOp::MIN)
{
return (ck::NumericLimits<AccDataType>::Max());
}
else if constexpr(ReduceOpId == ReduceTensorOp::MAX)
{
return (ck::NumericLimits<AccDataType>::Lowest());
}
else if constexpr(ReduceOpId == ReduceTensorOp::AMAX)
{
return (static_cast<AccDataType>(0.0f));
}
else
{
// ReduceTensorOp::ADD
// ReduceTensorOp::AVG
// ReduceTensorOp::NORM1
// ReduceTensorOp::NORM2
return (static_cast<AccDataType>(0.0f));
};
};
template <typename AccDataType, bool PropagateNan>
__host__ static inline void
binop_with_nan_check(std::function<void(AccDataType&, AccDataType)> opReduce,
AccDataType& accuVal,
AccDataType currVal)
{
using ck::math::isnan;
if constexpr(!PropagateNan)
{
opReduce(accuVal, currVal);
}
else
{
if(isnan(currVal))
accuVal = currVal;
else
opReduce(accuVal, currVal);
};
};
template <typename AccDataType, typename IndexDataType, bool PropagateNan>
__host__ static inline void
binop_with_index_and_nan_check(std::function<void(AccDataType&, AccDataType, bool&)> opReduce,
AccDataType& accuVal,
AccDataType currVal,
IndexDataType& accuIndex,
IndexDataType currIndex)
{
using ck::math::isnan;
if constexpr(!PropagateNan)
{
bool changed;
opReduce(accuVal, currVal, changed);
if(changed)
accuIndex = currIndex;
}
else
{
if(isnan(currVal))
{
accuVal = currVal;
accuIndex = currIndex;
}
else
{
bool changed;
opReduce(accuVal, currVal, changed);
if(changed)
accuIndex = currIndex;
};
};
};
}; // namespace host_reduce
}; // namespace ck
#endif

71
library/include/ck/library/host_tensor/host_reduction.hpp
View File

@@ -33,10 +33,10 @@
#include "reduction_enums.hpp"
#include "reduction_common.hpp"
#include "host_reduce_util.hpp"
#include "host_common_util.hpp"
#include "host_tensor.hpp"
#include "data_type.hpp"
#include "reduction_functions_accumulate.hpp"
template <int NDim>
static void get_all_indexes(const std::array<size_t, NDim>& dimLengths,
@@ -106,11 +106,13 @@ static size_t get_offset_from_index(const std::vector<size_t>& strides,
template <typename InDataType,
typename AccDataType,
typename OutDataType,
ck::ReduceTensorOp ReduceOpId,
typename ReduceOperation,
typename InElementwiseOperation,
typename AccElementwiseOperation,
int Rank,
int NumReduceDim,
bool PropagateNan,
bool NeedIndices>
bool OutputIndex>
struct ReductionHost
{
using IndexDataType = int32_t;
@@ -122,8 +124,6 @@ struct ReductionHost
std::vector<int> reduceDims;
IndexDataType divider;
std::function<void(AccDataType&)> preUnaryOp;
std::function<void(AccDataType&)> posUnaryOp;
std::array<size_t, NumReduceDim> reduceLengths;
std::array<size_t, NumReduceDim> reduceStrides;
std::array<size_t, NumInvariantDim> invariantLengths;
@@ -137,9 +137,6 @@ struct ReductionHost
const std::vector<int>& invariantDims_,
const std::vector<int>& reduceDims_)
{
using ck::host_reduce::PosUnaryOpFn;
using ck::host_reduce::PreUnaryOpFn;
// this->outLengths = to_int_vector(outDesc.GetLengths());
this->outStrides = outDesc.GetStrides();
@@ -171,9 +168,6 @@ struct ReductionHost
invariant_dim_indexes.clear();
get_all_indexes<NumInvariantDim>(invariantLengths, invariant_dim_indexes);
};
preUnaryOp = PreUnaryOpFn<AccDataType, ReduceOpId>(divider);
posUnaryOp = PosUnaryOpFn<AccDataType, ReduceOpId>(divider);
};
void Run(float alpha,
@@ -182,7 +176,7 @@ struct ReductionHost
OutDataType* out_data,
IndexDataType* out_indices)
{
if constexpr(NeedIndices)
if constexpr(OutputIndex)
{
RunImpl_with_index(alpha, in_data, beta, out_data, out_indices);
}
@@ -201,15 +195,17 @@ struct ReductionHost
using ck::float_equal_one;
using ck::float_equal_zero;
using ck::type_convert;
using ck::host_reduce::binop_with_index_and_nan_check;
using ck::host_reduce::ReduceOpFn2;
using ck::host_reduce::ReduceOpZeroVal;
auto opReduce2 = ReduceOpFn2<AccDataType, ReduceOpId>();
using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
ReduceOperation,
AccDataType,
IndexDataType>;
InElementwiseOperation in_elementwise_op(divider);
AccElementwiseOperation acc_elementwise_op(divider);
if constexpr(NumInvariantDim == 0)
{
AccDataType accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
AccDataType accuVal = ReduceOperation::GetIdentityValue();
IndexDataType accuIndex = 0;
for(std::size_t i = 0; i < reduce_dim_indexes.size(); i++)
@@ -219,15 +215,14 @@ struct ReductionHost
auto currVal = type_convert<AccDataType>(in_data[offset_reduce]);
preUnaryOp(currVal);
in_elementwise_op(currVal, currVal);
auto currIndex = static_cast<IndexDataType>(i);
binop_with_index_and_nan_check<AccDataType, IndexDataType, PropagateNan>(
opReduce2, accuVal, currVal, accuIndex, currIndex);
Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
};
posUnaryOp(accuVal);
acc_elementwise_op(accuVal, accuVal);
if(!float_equal_one{}(alpha))
accuVal *= type_convert<AccDataType>(alpha);
@@ -241,7 +236,7 @@ struct ReductionHost
else
{
auto thread_reduce_func = [&](auto invariant_index) {
AccDataType accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
AccDataType accuVal = ReduceOperation::GetIdentityValue();
IndexDataType accuIndex = 0;
auto offset_invariant =
@@ -255,15 +250,14 @@ struct ReductionHost
auto currVal =
type_convert<AccDataType>(in_data[offset_invariant + offset_reduce]);
preUnaryOp(currVal);
in_elementwise_op(currVal, currVal);
auto currIndex = static_cast<IndexDataType>(i);
binop_with_index_and_nan_check<AccDataType, IndexDataType, PropagateNan>(
opReduce2, accuVal, currVal, accuIndex, currIndex);
Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
};
posUnaryOp(accuVal);
acc_elementwise_op(accuVal, accuVal);
if(!float_equal_one{}(alpha))
accuVal *= type_convert<AccDataType>(alpha);
@@ -308,15 +302,16 @@ struct ReductionHost
using ck::float_equal_one;
using ck::float_equal_zero;
using ck::type_convert;
using ck::host_reduce::binop_with_nan_check;
using ck::host_reduce::ReduceOpFn;
using ck::host_reduce::ReduceOpZeroVal;
auto opReduce = ReduceOpFn<AccDataType, ReduceOpId>();
using Accumulation =
ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
InElementwiseOperation in_elementwise_op(divider);
AccElementwiseOperation acc_elementwise_op(divider);
if constexpr(NumInvariantDim == 0)
{
AccDataType accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
AccDataType accuVal = ReduceOperation::GetIdentityValue();
for(const auto& reduce_index : reduce_dim_indexes)
{
@@ -325,12 +320,12 @@ struct ReductionHost
auto currVal = type_convert<AccDataType>(in_data[offset_reduce]);
preUnaryOp(currVal);
in_elementwise_op(currVal, currVal);
binop_with_nan_check<AccDataType, PropagateNan>(opReduce, accuVal, currVal);
Accumulation::Calculate(accuVal, currVal);
};
posUnaryOp(accuVal);
acc_elementwise_op(accuVal, accuVal);
if(!float_equal_one{}(alpha))
accuVal *= type_convert<AccDataType>(alpha);
@@ -343,7 +338,7 @@ struct ReductionHost
else
{
auto thread_reduce_func = [&](auto invariant_index) {
AccDataType accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
AccDataType accuVal = ReduceOperation::GetIdentityValue();
auto offset_invariant =
get_offset_from_index<NumInvariantDim>(invariantStrides, invariant_index);
@@ -356,12 +351,12 @@ struct ReductionHost
auto currVal =
type_convert<AccDataType>(in_data[offset_invariant + offset_reduce]);
preUnaryOp(currVal);
in_elementwise_op(currVal, currVal);
binop_with_nan_check<AccDataType, PropagateNan>(opReduce, accuVal, currVal);
Accumulation::Calculate(accuVal, currVal);
};
posUnaryOp(accuVal);
acc_elementwise_op(accuVal, accuVal);
if(!float_equal_one{}(alpha))
accuVal *= type_convert<AccDataType>(alpha);