| 
 | 1 | +//! This benchmarks the `Integer::isqrt` methods.  | 
 | 2 | +
  | 
 | 3 | +macro_rules! benches {  | 
 | 4 | +    ($($T:ident)+) => {  | 
 | 5 | +        $(  | 
 | 6 | +            mod $T {  | 
 | 7 | +                use test::{black_box, Bencher};  | 
 | 8 | + | 
 | 9 | +                // Benchmark the square roots of:  | 
 | 10 | +                //  | 
 | 11 | +                // * the first 1,024 perfect squares  | 
 | 12 | +                // * halfway between each of the first 1,024 perfect squares  | 
 | 13 | +                //   and the next perfect square  | 
 | 14 | +                // * the next perfect square after the each of the first 1,024  | 
 | 15 | +                //   perfect squares, minus one  | 
 | 16 | +                // * the last 1,024 perfect squares  | 
 | 17 | +                // * the last 1,024 perfect squares, minus one  | 
 | 18 | +                // * halfway between each of the last 1,024 perfect squares  | 
 | 19 | +                //   and the previous perfect square  | 
 | 20 | +                #[bench]  | 
 | 21 | +                fn isqrt(bench: &mut Bencher) {  | 
 | 22 | +                    let mut inputs = Vec::with_capacity(6 * 1_024);  | 
 | 23 | + | 
 | 24 | +                    // The inputs to benchmark are worked out by using the fact  | 
 | 25 | +                    // that the nth nonzero perfect square is the sum of the  | 
 | 26 | +                    // first n odd numbers:  | 
 | 27 | +                    //  | 
 | 28 | +                    //  1 = 1  | 
 | 29 | +                    //  4 = 1 + 3  | 
 | 30 | +                    //  9 = 1 + 3 + 5  | 
 | 31 | +                    // 16 = 1 + 3 + 5 + 7  | 
 | 32 | +                    //  | 
 | 33 | +                    // Note also that the last odd number added in is two times  | 
 | 34 | +                    // the square root of the previous perfect square, plus  | 
 | 35 | +                    // one:  | 
 | 36 | +                    //  | 
 | 37 | +                    // 1 = 2*0 + 1  | 
 | 38 | +                    // 3 = 2*1 + 1  | 
 | 39 | +                    // 5 = 2*2 + 1  | 
 | 40 | +                    // 7 = 2*3 + 1  | 
 | 41 | +                    //  | 
 | 42 | +                    // That means we can add the square root of this perfect  | 
 | 43 | +                    // square once to get about halfway to the next perfect  | 
 | 44 | +                    // square, then we can add the square root of this perfect  | 
 | 45 | +                    // square again to get to the next perfect square minus  | 
 | 46 | +                    // one, then we can add one to get to the next perfect  | 
 | 47 | +                    // square.  | 
 | 48 | +                    //  | 
 | 49 | +                    // Here we include, for each of the first 1,024 perfect  | 
 | 50 | +                    // squares:  | 
 | 51 | +                    //  | 
 | 52 | +                    // * the current perfect square  | 
 | 53 | +                    // * about halfway to the next perfect square  | 
 | 54 | +                    // * the next perfect square, minus one  | 
 | 55 | +                    let mut n: $T = 0;  | 
 | 56 | +                    for sqrt_n in 0..1_024.min((1_u128 << (($T::BITS - $T::MAX.leading_zeros())/2)) - 1) as $T {  | 
 | 57 | +                        inputs.push(n);  | 
 | 58 | +                        n += sqrt_n;  | 
 | 59 | +                        inputs.push(n);  | 
 | 60 | +                        n += sqrt_n;  | 
 | 61 | +                        inputs.push(n);  | 
 | 62 | +                        n += 1;  | 
 | 63 | +                    }  | 
 | 64 | + | 
 | 65 | +                    // Similarly, we include, for each of the last 1,024  | 
 | 66 | +                    // perfect squares:  | 
 | 67 | +                    //  | 
 | 68 | +                    // * the current perfect square  | 
 | 69 | +                    // * the current perfect square, minus one  | 
 | 70 | +                    // * about halfway to the previous perfect square  | 
 | 71 | +                    let maximum_sqrt = $T::MAX.isqrt();  | 
 | 72 | +                    let mut n = maximum_sqrt * maximum_sqrt;  | 
 | 73 | + | 
 | 74 | +                    for sqrt_n in (maximum_sqrt - 1_024.min((1_u128 << (($T::BITS - 1)/2)) - 1) as $T..maximum_sqrt).rev() {  | 
 | 75 | +                        inputs.push(n);  | 
 | 76 | +                        n -= 1;  | 
 | 77 | +                        inputs.push(n);  | 
 | 78 | +                        n -= sqrt_n;  | 
 | 79 | +                        inputs.push(n);  | 
 | 80 | +                        n -= sqrt_n;  | 
 | 81 | +                    }  | 
 | 82 | + | 
 | 83 | +                    bench.iter(|| {  | 
 | 84 | +                        for x in &inputs {  | 
 | 85 | +                            black_box(black_box(x).isqrt());  | 
 | 86 | +                        }  | 
 | 87 | +                    });  | 
 | 88 | +                }  | 
 | 89 | +            }  | 
 | 90 | +        )*  | 
 | 91 | +    };  | 
 | 92 | +}  | 
 | 93 | + | 
 | 94 | +macro_rules! push_n {  | 
 | 95 | +    ($T:ident, $inputs:ident, $n:ident) => {  | 
 | 96 | +        if $n != 0 {  | 
 | 97 | +            $inputs.push(  | 
 | 98 | +                core::num::$T::new($n)  | 
 | 99 | +                    .expect("Cannot create a new `NonZero` value from a nonzero value"),  | 
 | 100 | +            );  | 
 | 101 | +        }  | 
 | 102 | +    };  | 
 | 103 | +}  | 
 | 104 | + | 
 | 105 | +macro_rules! nonzero_benches {  | 
 | 106 | +    ($mod:ident $T:ident $RegularT:ident) => {  | 
 | 107 | +        mod $mod {  | 
 | 108 | +            use test::{black_box, Bencher};  | 
 | 109 | + | 
 | 110 | +            // Benchmark the square roots of:  | 
 | 111 | +            //  | 
 | 112 | +            // * the first 1,024 perfect squares  | 
 | 113 | +            // * halfway between each of the first 1,024 perfect squares  | 
 | 114 | +            //   and the next perfect square  | 
 | 115 | +            // * the next perfect square after the each of the first 1,024  | 
 | 116 | +            //   perfect squares, minus one  | 
 | 117 | +            // * the last 1,024 perfect squares  | 
 | 118 | +            // * the last 1,024 perfect squares, minus one  | 
 | 119 | +            // * halfway between each of the last 1,024 perfect squares  | 
 | 120 | +            //   and the previous perfect square  | 
 | 121 | +            #[bench]  | 
 | 122 | +            fn isqrt(bench: &mut Bencher) {  | 
 | 123 | +                let mut inputs: Vec<core::num::$T> = Vec::with_capacity(6 * 1_024);  | 
 | 124 | + | 
 | 125 | +                // The inputs to benchmark are worked out by using the fact  | 
 | 126 | +                // that the nth nonzero perfect square is the sum of the  | 
 | 127 | +                // first n odd numbers:  | 
 | 128 | +                //  | 
 | 129 | +                //  1 = 1  | 
 | 130 | +                //  4 = 1 + 3  | 
 | 131 | +                //  9 = 1 + 3 + 5  | 
 | 132 | +                // 16 = 1 + 3 + 5 + 7  | 
 | 133 | +                //  | 
 | 134 | +                // Note also that the last odd number added in is two times  | 
 | 135 | +                // the square root of the previous perfect square, plus  | 
 | 136 | +                // one:  | 
 | 137 | +                //  | 
 | 138 | +                // 1 = 2*0 + 1  | 
 | 139 | +                // 3 = 2*1 + 1  | 
 | 140 | +                // 5 = 2*2 + 1  | 
 | 141 | +                // 7 = 2*3 + 1  | 
 | 142 | +                //  | 
 | 143 | +                // That means we can add the square root of this perfect  | 
 | 144 | +                // square once to get about halfway to the next perfect  | 
 | 145 | +                // square, then we can add the square root of this perfect  | 
 | 146 | +                // square again to get to the next perfect square minus  | 
 | 147 | +                // one, then we can add one to get to the next perfect  | 
 | 148 | +                // square.  | 
 | 149 | +                //  | 
 | 150 | +                // Here we include, for each of the first 1,024 perfect  | 
 | 151 | +                // squares:  | 
 | 152 | +                //  | 
 | 153 | +                // * the current perfect square  | 
 | 154 | +                // * about halfway to the next perfect square  | 
 | 155 | +                // * the next perfect square, minus one  | 
 | 156 | +                let mut n: $RegularT = 0;  | 
 | 157 | +                for sqrt_n in 0..1_024  | 
 | 158 | +                    .min((1_u128 << (($RegularT::BITS - $RegularT::MAX.leading_zeros()) / 2)) - 1)  | 
 | 159 | +                    as $RegularT  | 
 | 160 | +                {  | 
 | 161 | +                    push_n!($T, inputs, n);  | 
 | 162 | +                    n += sqrt_n;  | 
 | 163 | +                    push_n!($T, inputs, n);  | 
 | 164 | +                    n += sqrt_n;  | 
 | 165 | +                    push_n!($T, inputs, n);  | 
 | 166 | +                    n += 1;  | 
 | 167 | +                }  | 
 | 168 | + | 
 | 169 | +                // Similarly, we include, for each of the last 1,024  | 
 | 170 | +                // perfect squares:  | 
 | 171 | +                //  | 
 | 172 | +                // * the current perfect square  | 
 | 173 | +                // * the current perfect square, minus one  | 
 | 174 | +                // * about halfway to the previous perfect square  | 
 | 175 | +                let maximum_sqrt = $RegularT::MAX.isqrt();  | 
 | 176 | +                let mut n = maximum_sqrt * maximum_sqrt;  | 
 | 177 | + | 
 | 178 | +                for sqrt_n in (maximum_sqrt  | 
 | 179 | +                    - 1_024.min((1_u128 << (($RegularT::BITS - 1) / 2)) - 1) as $RegularT  | 
 | 180 | +                    ..maximum_sqrt)  | 
 | 181 | +                    .rev()  | 
 | 182 | +                {  | 
 | 183 | +                    push_n!($T, inputs, n);  | 
 | 184 | +                    n -= 1;  | 
 | 185 | +                    push_n!($T, inputs, n);  | 
 | 186 | +                    n -= sqrt_n;  | 
 | 187 | +                    push_n!($T, inputs, n);  | 
 | 188 | +                    n -= sqrt_n;  | 
 | 189 | +                }  | 
 | 190 | + | 
 | 191 | +                bench.iter(|| {  | 
 | 192 | +                    for n in &inputs {  | 
 | 193 | +                        black_box(black_box(n).isqrt());  | 
 | 194 | +                    }  | 
 | 195 | +                });  | 
 | 196 | +            }  | 
 | 197 | +        }  | 
 | 198 | +    };  | 
 | 199 | +}  | 
 | 200 | + | 
 | 201 | +benches!(i8 i16 i32 i64 i128 isize u8 u16 u32 u64 u128 usize);  | 
 | 202 | +nonzero_benches!(non_zero_u8 NonZeroU8 u8);  | 
 | 203 | +nonzero_benches!(non_zero_u16 NonZeroU16 u16);  | 
 | 204 | +nonzero_benches!(non_zero_u32 NonZeroU32 u32);  | 
 | 205 | +nonzero_benches!(non_zero_u64 NonZeroU64 u64);  | 
 | 206 | +nonzero_benches!(non_zero_u128 NonZeroU128 u128);  | 
 | 207 | +nonzero_benches!(non_zero_usize NonZeroUsize usize);  | 
0 commit comments