runtime dispatching (fixed)

ydb/library/yql/utils/simd/exec/runtime_dispatching/AVX2_algo/avx2_algo.cpp

@@ -0,0 +1,9 @@
+#include <ydb/library/yql/utils/simd/exec/runtime_dispatching/algo.h>
+#include <immintrin.h>
+#include <avxintrin.h>
+
+template<>
+THolder<Perfomancer::Interface> Perfomancer::Create<AVX2Trait>() {
+    Cerr << "AVXTrait ";
+    return MakeHolder<Algo<AVX2Trait>>();
+}

ydb/library/yql/utils/simd/exec/runtime_dispatching/AVX2_algo/ya.make

@@ -0,0 +1,9 @@
+LIBRARY()
+
+OWNER(g:yql)
+
+CFLAGS(-mavx2)
+
+SRCS(avx2_algo.cpp)
+
+END()

ydb/library/yql/utils/simd/exec/runtime_dispatching/Fallback_algo/fallback_algo.cpp

@@ -0,0 +1,7 @@
+#include <ydb/library/yql/utils/simd/exec/runtime_dispatching/algo.h>
+
+template<>
+THolder<Perfomancer::Interface> Perfomancer::Create<FallbackTrait>() {
+    Cerr << "FallbackTrait ";
+    return MakeHolder<Algo<FallbackTrait>>();
+}

ydb/library/yql/utils/simd/exec/runtime_dispatching/Fallback_algo/ya.make

@@ -0,0 +1,7 @@
+LIBRARY()
+
+OWNER(g:yql)
+
+SRCS(fallback_algo.cpp)
+
+END()

ydb/library/yql/utils/simd/exec/runtime_dispatching/SSE42_algo/sse42_algo.cpp

@@ -0,0 +1,8 @@
+#include <ydb/library/yql/utils/simd/exec/runtime_dispatching/algo.h>
+#include <immintrin.h>
+
+template<>
+THolder<Perfomancer::Interface> Perfomancer::Create<SSE42Trait>() {
+    Cerr << "SSETrait ";
+    return MakeHolder<Algo<SSE42Trait>>();
+}

ydb/library/yql/utils/simd/exec/runtime_dispatching/SSE42_algo/ya.make

@@ -0,0 +1,9 @@
+LIBRARY()
+
+OWNER(g:yql)
+
+CFLAGS(-msse4.2)
+
+SRCS(sse42_algo.cpp)
+
+END()

ydb/library/yql/utils/simd/exec/runtime_dispatching/algo.h

@@ -0,0 +1,108 @@
+#include "util/system/cpu_id.h"
+#include <ydb/library/yql/utils/simd/simd_avx2.h>
+#include <ydb/library/yql/utils/simd/simd_fallback.h>
+#include <ydb/library/yql/utils/simd/simd_sse42.h>
+#include <vector>
+
+using vl = std::vector<ui64>;
+using vvl = std::vector<std::vector<ui64>>;
+
+using AVX2Trait = NSimd::NAVX2::TSimd8<ui64>;
+
+using SSE42Trait = NSimd::NSSE42::TSimd8<ui64>;
+
+using FallbackTrait = NSimd::NFallback::FallbackTrait<ui64>;
+
+struct Perfomancer {
+
+    Perfomancer() = default;
+
+    struct Interface {
+
+        virtual ~Interface() = default;
+
+        inline virtual void Add(vvl& columns, vl& result) {
+
+            // to avoid clang(-Wunused-parameter)
+            columns[0];
+            result[0];
+        }
+
+    };
+
+
+    template <typename Trait>
+    struct Algo : Interface {
+
+        Algo() {}
+
+        inline void Add(vvl& columns, vl& result) override {
+            std::chrono::steady_clock::time_point begin = std::chrono::steady_clock::now();
+
+            std::vector<Trait> Registers(columns.size());
+
+            for (size_t j = 0; j < result.size(); j += Trait::SIZE / sizeof(ui64)) {
+
+                for (size_t i = 0; i < columns.size(); ++i) {
+                    Registers[i] = Trait(&columns[i][j]);
+                }
+
+                for (size_t i = 1; i < columns.size(); ++i) {
+                    Registers[i] += Registers[i - 1];
+                }
+
+                Registers.back().Store(&result[j]);
+            }
+
+            Cerr << std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::steady_clock::now() - begin).count() << "ms\n";
+        }
+
+        ~Algo() = default;
+    };
+
+    template <typename Trait>
+    inline THolder<Interface> Create() {
+        return MakeHolder<Interface>();
+    }
+
+};
+
+template<>
+THolder<Perfomancer::Interface> Perfomancer::Create<AVX2Trait>();
+
+template<>
+THolder<Perfomancer::Interface> Perfomancer::Create<SSE42Trait>();
+
+template<>
+THolder<Perfomancer::Interface> Perfomancer::Create<FallbackTrait>();
+
+template <typename TFactory>
+auto ChooseTrait(TFactory& factory) {
+
+    if (NX86::HaveAVX2()) {
+        return factory.template Create<AVX2Trait>();
+
+    } else if (NX86::HaveSSE42()) {
+        return factory.template Create<SSE42Trait>();
+
+    }
+
+    return factory.template Create<FallbackTrait>();
+}
+
+//this part of code just to compare times of work
+//we dont need this functions at all
+template <typename TFactory>
+auto ChooseAVX2Trait(TFactory& factory) {
+    return factory.template Create<AVX2Trait>();
+}
+
+template <typename TFactory>
+auto ChooseSSE42Trait(TFactory& factory) {
+    return factory.template Create<SSE42Trait>();
+}
+
+template <typename TFactory>
+auto ChooseFallbackTrait(TFactory& factory) {
+    return factory.template Create<FallbackTrait>();
+}

ydb/library/yql/utils/simd/exec/runtime_dispatching/main/main.cpp

@@ -0,0 +1,30 @@
+#include <ydb/library/yql/utils/simd/exec/runtime_dispatching/algo.h>
+
+int main() {
+
+    std::vector<std::vector<ui64>> columns(10, std::vector<ui64>(64e6, 1e12 + 7));
+    std::vector<ui64> result(64e6);
+
+    Perfomancer perfomancer;
+
+    Cerr << "Best Trait is: ";
+    auto worker = ChooseTrait(perfomancer);
+
+    worker->Add(columns, result);
+    result.assign(64e6, 0);
+
+    if (NX86::HaveSSE42()) {
+        auto SSE42worker = ChooseSSE42Trait(perfomancer);
+        SSE42worker->Add(columns, result);
+        result.assign(64e6, 0);
+    }
+
+    if (NX86::HaveAVX2()) {
+        auto AVX2worker = ChooseAVX2Trait(perfomancer);
+        AVX2worker->Add(columns, result);
+        result.assign(64e6, 0);
+    }
+
+    auto Fallbackworker = ChooseFallbackTrait(perfomancer);
+    Fallbackworker->Add(columns, result);
+}

ydb/library/yql/utils/simd/exec/runtime_dispatching/main/ya.make

@@ -0,0 +1,9 @@
+OWNER(g:yql)
+
+PROGRAM()
+
+SRCS(main.cpp)
+
+PEERDIR(ydb/library/yql/utils/simd/exec/runtime_dispatching)
+
+END()

ydb/library/yql/utils/simd/exec/runtime_dispatching/ya.make

@@ -0,0 +1,18 @@
+OWNER(g:yql)
+
+LIBRARY()
+
+PEERDIR(
+    ydb/library/yql/utils/simd/exec/runtime_dispatching/AVX2_algo
+    ydb/library/yql/utils/simd/exec/runtime_dispatching/SSE42_algo
+    ydb/library/yql/utils/simd/exec/runtime_dispatching/Fallback_algo
+)
+
+END()
+
+RECURSE(
+    AVX2_algo
+    SSE42_algo
+    Fallback_algo
+    main
+)

ydb/library/yql/utils/simd/simd_avx2.h

@@ -426,6 +426,17 @@ struct TSimd8 {
     }
 };
 
+template<>
+inline TSimd8<ui64> TSimd8<ui64>::operator+(const TSimd8<ui64>& other) const {
+    return _mm256_add_epi64(Value, other.Value);
+}
+
+template<>
+inline TSimd8<ui64>& TSimd8<ui64>::operator+=(const TSimd8<ui64>& other) {
+    *this = *this + other.Value;
+    return *this;
+}
+
 template<>
 inline TSimd8<bool> TSimd8<bool>::Set(bool value) {
     return _mm256_set1_epi8(ui8(-value));

ydb/library/yql/utils/simd/simd_fallback.h

@@ -63,6 +63,44 @@ struct TBase {
     };
 };
 
+template <typename T>
+struct FallbackTrait {
+
+    T Value;
+
+    static const int SIZE = sizeof(T);
+
+    inline FallbackTrait() : Value() {}
+
+    inline FallbackTrait(const FallbackTrait& other) : Value(other.Value) {}
+
+    inline FallbackTrait(const T* ptr) : Value(*ptr) {}
+
+    inline FallbackTrait& operator=(const FallbackTrait& other) {
+        if (&other == this) return *this;
+
+        Value = other.Value;
+        return *this;
+    }
+
+    inline FallbackTrait& operator+=(const FallbackTrait& other) {
+        Value += other.Value;
+        return *this;
+    }
+
+    inline FallbackTrait operator+(const FallbackTrait& other) {
+        FallbackTrait ans;
+
+        ans += other;
+
+        return ans;
+    }
+
+    inline void Store(T* ptr) {
+        *ptr = Value;
+    }
+};
+
 template<typename T, typename Mask=TSimd8<bool>>
 struct TBase8: TBase<TSimd8<T>> {
 

ydb/library/yql/utils/simd/simd_sse42.h

@@ -15,7 +15,7 @@ template<typename T>
 struct TSimd8 {
     __m128i Value;
 
-    static const int SIZE = 32;
+    static const int SIZE = 16;
 
     inline TSimd8()
         : Value{__m128i()} {
@@ -339,6 +339,17 @@ struct TSimd8 {
     }
 };
 
+template<>
+inline TSimd8<ui64> TSimd8<ui64>::operator+(const TSimd8<ui64>& other) const {
+    return _mm_add_epi64(Value, other.Value);
+}
+
+template<>
+inline TSimd8<ui64>& TSimd8<ui64>::operator+=(const TSimd8<ui64>& other) {
+    *this = *this + other.Value;
+    return *this;
+}
+
 template<>
 inline TSimd8<bool> TSimd8<bool>::Set(bool value) {
     return _mm_set1_epi8(ui8(-value));