| 1 | #include <Functions/FunctionFactory.h> |
| 2 | #include <Functions/FunctionBinaryArithmetic.h> |
| 3 | #include <Functions/intDiv.h> |
| 4 | |
| 5 | #ifdef __SSE2__ |
| 6 | #define LIBDIVIDE_USE_SSE2 1 |
| 7 | #endif |
| 8 | |
| 9 | #include <libdivide.h> |
| 10 | |
| 11 | |
| 12 | namespace DB |
| 13 | { |
| 14 | |
| 15 | /// Optimizations for integer division by a constant. |
| 16 | |
| 17 | template <typename A, typename B> |
| 18 | struct DivideIntegralByConstantImpl |
| 19 | : BinaryOperationImplBase<A, B, DivideIntegralImpl<A, B>> |
| 20 | { |
| 21 | using ResultType = typename DivideIntegralImpl<A, B>::ResultType; |
| 22 | |
| 23 | static void vector_constant(const PaddedPODArray<A> & a, B b, PaddedPODArray<ResultType> & c) |
| 24 | { |
| 25 | if (unlikely(b == 0)) |
| 26 | throw Exception("Division by zero" , ErrorCodes::ILLEGAL_DIVISION); |
| 27 | |
| 28 | #pragma GCC diagnostic push |
| 29 | #pragma GCC diagnostic ignored "-Wsign-compare" |
| 30 | |
| 31 | if (unlikely(is_signed_v<B> && b == -1)) |
| 32 | { |
| 33 | size_t size = a.size(); |
| 34 | for (size_t i = 0; i < size; ++i) |
| 35 | c[i] = -c[i]; |
| 36 | return; |
| 37 | } |
| 38 | |
| 39 | #pragma GCC diagnostic pop |
| 40 | |
| 41 | libdivide::divider<A> divider(b); |
| 42 | |
| 43 | size_t size = a.size(); |
| 44 | const A * a_pos = a.data(); |
| 45 | const A * a_end = a_pos + size; |
| 46 | ResultType * c_pos = c.data(); |
| 47 | |
| 48 | #ifdef __SSE2__ |
| 49 | static constexpr size_t values_per_sse_register = 16 / sizeof(A); |
| 50 | const A * a_end_sse = a_pos + size / values_per_sse_register * values_per_sse_register; |
| 51 | |
| 52 | while (a_pos < a_end_sse) |
| 53 | { |
| 54 | _mm_storeu_si128(reinterpret_cast<__m128i *>(c_pos), |
| 55 | _mm_loadu_si128(reinterpret_cast<const __m128i *>(a_pos)) / divider); |
| 56 | |
| 57 | a_pos += values_per_sse_register; |
| 58 | c_pos += values_per_sse_register; |
| 59 | } |
| 60 | #endif |
| 61 | |
| 62 | while (a_pos < a_end) |
| 63 | { |
| 64 | *c_pos = *a_pos / divider; |
| 65 | ++a_pos; |
| 66 | ++c_pos; |
| 67 | } |
| 68 | } |
| 69 | }; |
| 70 | |
| 71 | /** Specializations are specified for dividing numbers of the type UInt64 and UInt32 by the numbers of the same sign. |
| 72 | * Can be expanded to all possible combinations, but more code is needed. |
| 73 | */ |
| 74 | |
| 75 | template <> struct BinaryOperationImpl<UInt64, UInt8, DivideIntegralImpl<UInt64, UInt8>> : DivideIntegralByConstantImpl<UInt64, UInt8> {}; |
| 76 | template <> struct BinaryOperationImpl<UInt64, UInt16, DivideIntegralImpl<UInt64, UInt16>> : DivideIntegralByConstantImpl<UInt64, UInt16> {}; |
| 77 | template <> struct BinaryOperationImpl<UInt64, UInt32, DivideIntegralImpl<UInt64, UInt32>> : DivideIntegralByConstantImpl<UInt64, UInt32> {}; |
| 78 | template <> struct BinaryOperationImpl<UInt64, UInt64, DivideIntegralImpl<UInt64, UInt64>> : DivideIntegralByConstantImpl<UInt64, UInt64> {}; |
| 79 | |
| 80 | template <> struct BinaryOperationImpl<UInt32, UInt8, DivideIntegralImpl<UInt32, UInt8>> : DivideIntegralByConstantImpl<UInt32, UInt8> {}; |
| 81 | template <> struct BinaryOperationImpl<UInt32, UInt16, DivideIntegralImpl<UInt32, UInt16>> : DivideIntegralByConstantImpl<UInt32, UInt16> {}; |
| 82 | template <> struct BinaryOperationImpl<UInt32, UInt32, DivideIntegralImpl<UInt32, UInt32>> : DivideIntegralByConstantImpl<UInt32, UInt32> {}; |
| 83 | template <> struct BinaryOperationImpl<UInt32, UInt64, DivideIntegralImpl<UInt32, UInt64>> : DivideIntegralByConstantImpl<UInt32, UInt64> {}; |
| 84 | |
| 85 | template <> struct BinaryOperationImpl<Int64, Int8, DivideIntegralImpl<Int64, Int8>> : DivideIntegralByConstantImpl<Int64, Int8> {}; |
| 86 | template <> struct BinaryOperationImpl<Int64, Int16, DivideIntegralImpl<Int64, Int16>> : DivideIntegralByConstantImpl<Int64, Int16> {}; |
| 87 | template <> struct BinaryOperationImpl<Int64, Int32, DivideIntegralImpl<Int64, Int32>> : DivideIntegralByConstantImpl<Int64, Int32> {}; |
| 88 | template <> struct BinaryOperationImpl<Int64, Int64, DivideIntegralImpl<Int64, Int64>> : DivideIntegralByConstantImpl<Int64, Int64> {}; |
| 89 | |
| 90 | template <> struct BinaryOperationImpl<Int32, Int8, DivideIntegralImpl<Int32, Int8>> : DivideIntegralByConstantImpl<Int32, Int8> {}; |
| 91 | template <> struct BinaryOperationImpl<Int32, Int16, DivideIntegralImpl<Int32, Int16>> : DivideIntegralByConstantImpl<Int32, Int16> {}; |
| 92 | template <> struct BinaryOperationImpl<Int32, Int32, DivideIntegralImpl<Int32, Int32>> : DivideIntegralByConstantImpl<Int32, Int32> {}; |
| 93 | template <> struct BinaryOperationImpl<Int32, Int64, DivideIntegralImpl<Int32, Int64>> : DivideIntegralByConstantImpl<Int32, Int64> {}; |
| 94 | |
| 95 | |
| 96 | struct NameIntDiv { static constexpr auto name = "intDiv" ; }; |
| 97 | using FunctionIntDiv = FunctionBinaryArithmetic<DivideIntegralImpl, NameIntDiv, false>; |
| 98 | |
| 99 | void registerFunctionIntDiv(FunctionFactory & factory) |
| 100 | { |
| 101 | factory.registerFunction<FunctionIntDiv>(); |
| 102 | } |
| 103 | |
| 104 | } |
| 105 | |