ReductionExpression.h

/*
  ==============================================================================
    DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
 
    Copyright 2022- by sonible GmbH.
 
    This file is part of VCTR - Versatile Container Templates Reconceptualized.
 
    VCTR is free software: you can redistribute it and/or modify
    it under the terms of the GNU Lesser General Public License version 3
    only, as published by the Free Software Foundation.
 
    VCTR is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    GNU Lesser General Public License version 3 for more details.
 
    You should have received a copy of the GNU Lesser General Public License
    version 3 along with VCTR.  If not, see <https://www.gnu.org/licenses/>.
  ==============================================================================
*/
 
namespace vctr
{
 
class ReductionExpression
{
public:
    //==============================================================================
 
    template <is::reductionExpression Expression>
    VCTR_FORCEDINLINE static constexpr auto reduce (const Expression& e)
    {
        const auto s = e.size();
 
        // Reducing an empty vector can lead to undefined edge cases
        VCTR_ASSERT (s > 0);
 
        if (! std::is_constant_evaluated())
        {
            if constexpr (has::reduceVectorOp<Expression, ValueType<Expression>>)
            {
                return e.reduceVectorOp();
            }
 
            if constexpr (has::reduceNeonRegisterWise<Expression, ValueType<Expression>>)
            {
                return reduceNeon (e);
            }
 
            if constexpr (has::reduceAVXRegisterWise<Expression, ValueType<Expression>>)
            {
                if constexpr (is::realFloatNumber<ValueType<Expression>>)
                {
                    if (Config::supportedCPUInstructionSets.fma)
                        return reduceFMA (e);
                }
                else
                {
                    if (Config::supportedCPUInstructionSets.avx2)
                        return reduceAVX2 (e);
                }
            }
 
            if constexpr (has::reduceSSERegisterWise<Expression, ValueType<Expression>>)
            {
                if (Config::supportedCPUInstructionSets.sse4_1)
                    return reduceSSE (e);
            }
        }
 
        std::array<ValueType<Expression>, 1> v = { Expression::reductionResultInitValue };
        for (size_t i = 0; i < s; ++i)
            e.reduceElementWise (v[0], i);
 
        return e.finalizeReduction (v);
    }
 
private:
    //==============================================================================
    template <is::reductionExpression Expression>
    VCTR_TARGET ("avx2") static auto reduceAVX2 (const Expression& e)
    requires Config::archX64
    {
        using VType = ValueType<Expression>;
        using RType = AVXRegister<VType>;
 
        constexpr auto inc = RType::numElements;
        const auto n = e.size();
        const auto nSIMD = detail::previousMultipleOf<inc> (n);
 
        e.prepareAVXEvaluation();
 
        auto avxValue = RType::broadcast (Expression::reductionResultInitValue);
 
        size_t i = 0;
        for (; i < nSIMD; i += inc)
            e.reduceAVXRegisterWise (avxValue, i);
 
        alignas (Config::maxSIMDRegisterSize) std::array<VType, RType::numElements + 1> results;
        avxValue.storeAligned (results.data());
        results.back() = Expression::reductionResultInitValue;
 
        for (; i < n; ++i)
            e.reduceElementWise (results.back(), i);
 
        return e.finalizeReduction (results);
    }
 
    template <is::reductionExpression Expression>
    VCTR_TARGET ("fma") static auto reduceFMA (const Expression& e)
    requires Config::archX64
    {
        using VType = ValueType<Expression>;
        using RType = AVXRegister<VType>;
 
        constexpr auto inc = RType::numElements;
        const auto n = e.size();
        const auto nSIMD = detail::previousMultipleOf<inc> (n);
 
        e.prepareAVXEvaluation();
 
        auto avxValue = RType::broadcast (Expression::reductionResultInitValue);
 
        size_t i = 0;
        for (; i < nSIMD; i += inc)
            e.reduceAVXRegisterWise (avxValue, i);
 
        alignas (Config::maxSIMDRegisterSize) std::array<VType, RType::numElements + 1> results;
        avxValue.storeAligned (results.data());
        results.back() = Expression::reductionResultInitValue;
 
        for (; i < n; ++i)
            e.reduceElementWise (results.back(), i);
 
        return e.finalizeReduction (results);
    }
 
    template <is::reductionExpression Expression>
    VCTR_TARGET ("sse4.1") static auto reduceSSE (const Expression& e)
    requires Config::archX64
    {
        using VType = ValueType<Expression>;
        using RType = SSERegister<VType>;
 
        constexpr auto inc = RType::numElements;
        const auto n = e.size();
        const auto nSIMD = detail::previousMultipleOf<inc> (n);
 
        e.prepareSSEEvaluation();
 
        auto sseValue = RType::broadcast (Expression::reductionResultInitValue);
 
        size_t i = 0;
        for (; i < nSIMD; i += inc)
            e.reduceSSERegisterWise (sseValue, i);
 
        alignas (Config::maxSIMDRegisterSize) std::array<VType, RType::numElements + 1> results;
        sseValue.storeAligned (results.data());
        results.back() = Expression::reductionResultInitValue;
 
        for (; i < n; ++i)
            e.reduceElementWise (results.back(), i);
 
        return e.finalizeReduction (results);
    }
 
    template <is::reductionExpression Expression>
    static auto reduceNeon (const Expression& e)
    requires Config::archARM
    {
        using VType = ValueType<Expression>;
        using RType = NeonRegister<VType>;
 
        constexpr auto inc = RType::numElements;
        const auto n = e.size();
        const auto nSIMD = detail::previousMultipleOf<inc> (n);
 
        e.prepareNeonEvaluation();
 
        auto neonValue = RType::broadcast (Expression::reductionResultInitValue);
 
        size_t i = 0;
        for (; i < nSIMD; i += inc)
            e.reduceNeonRegisterWise (neonValue, i);
 
        alignas (Config::maxSIMDRegisterSize) std::array<VType, RType::numElements + 1> results;
        neonValue.store (results.data());
        results.back() = Expression::reductionResultInitValue;
 
        for (; i < n; ++i)
            e.reduceElementWise (results.back(), i);
 
        return e.finalizeReduction (results);
    }
};
 
} // namespace vctr