zig/std/rand/index.zig

653 lines
18 KiB
Zig

// The engines provided here should be initialized from an external source. For now, getRandomBytes
// from the os package is the most suitable. Be sure to use a CSPRNG when required, otherwise using
// a normal PRNG will be faster and use substantially less stack space.
//
// ```
// var buf: [8]u8 = undefined;
// try std.os.getRandomBytes(buf[0..]);
// const seed = mem.readInt(buf[0..8], u64, builtin.Endian.Little);
//
// var r = DefaultPrng.init(seed);
//
// const s = r.random.scalar(u64);
// ```
//
// TODO(tiehuis): Benchmark these against other reference implementations.
const std = @import("../index.zig");
const builtin = @import("builtin");
const assert = std.debug.assert;
const mem = std.mem;
const math = std.math;
// When you need fast unbiased random numbers
pub const DefaultPrng = Xoroshiro128;
// When you need cryptographically secure random numbers
pub const DefaultCsprng = Isaac64;
pub const Random = struct {
fillFn: fn(r: &Random, buf: []u8) void,
/// Read random bytes into the specified buffer until fill.
pub fn bytes(r: &Random, buf: []u8) void {
r.fillFn(r, buf);
}
/// Return a random integer/boolean type.
pub fn scalar(r: &Random, comptime T: type) T {
var rand_bytes: [@sizeOf(T)]u8 = undefined;
r.bytes(rand_bytes[0..]);
if (T == bool) {
return rand_bytes[0] & 0b1 == 0;
} else {
// NOTE: Cannot @bitCast array to integer type.
return mem.readInt(rand_bytes, T, builtin.Endian.Little);
}
}
/// Get a random unsigned integer with even distribution between `start`
/// inclusive and `end` exclusive.
pub fn range(r: &Random, comptime T: type, start: T, end: T) T {
assert(start <= end);
if (T.is_signed) {
const uint = @IntType(false, T.bit_count);
if (start >= 0 and end >= 0) {
return T(r.range(uint, uint(start), uint(end)));
} else if (start < 0 and end < 0) {
// Can't overflow because the range is over signed ints
return math.negateCast(r.range(uint, math.absCast(end), math.absCast(start)) + 1) catch unreachable;
} else if (start < 0 and end >= 0) {
const end_uint = uint(end);
const total_range = math.absCast(start) + end_uint;
const value = r.range(uint, 0, total_range);
const result = if (value < end_uint) x: {
break :x T(value);
} else if (value == end_uint) x: {
break :x start;
} else x: {
// Can't overflow because the range is over signed ints
break :x math.negateCast(value - end_uint) catch unreachable;
};
return result;
} else {
unreachable;
}
} else {
const total_range = end - start;
const leftover = @maxValue(T) % total_range;
const upper_bound = @maxValue(T) - leftover;
var rand_val_array: [@sizeOf(T)]u8 = undefined;
while (true) {
r.bytes(rand_val_array[0..]);
const rand_val = mem.readInt(rand_val_array, T, builtin.Endian.Little);
if (rand_val < upper_bound) {
return start + (rand_val % total_range);
}
}
}
}
/// Return a floating point value evenly distributed in the range [0, 1).
pub fn float(r: &Random, comptime T: type) T {
// Generate a uniform value between [1, 2) and scale down to [0, 1).
// Note: The lowest mantissa bit is always set to 0 so we only use half the available range.
switch (T) {
f32 => {
const s = r.scalar(u32);
const repr = (0x7f << 23) | (s >> 9);
return @bitCast(f32, repr) - 1.0;
},
f64 => {
const s = r.scalar(u64);
const repr = (0x3ff << 52) | (s >> 12);
return @bitCast(f64, repr) - 1.0;
},
else => @compileError("unknown floating point type"),
}
}
/// Return a floating point value normally distributed in the range [0, 1].
pub fn floatNorm(r: &Random, comptime T: type) T {
// TODO(tiehuis): See https://www.doornik.com/research/ziggurat.pdf
@compileError("floatNorm is unimplemented");
}
/// Return a exponentially distributed float between (0, @maxValue(f64))
pub fn floatExp(r: &Random, comptime T: type) T {
@compileError("floatExp is unimplemented");
}
/// Shuffle a slice into a random order.
pub fn shuffle(r: &Random, comptime T: type, buf: []T) void {
if (buf.len < 2) {
return;
}
var i: usize = 0;
while (i < buf.len - 1) : (i += 1) {
const j = r.range(usize, i, buf.len);
mem.swap(T, &buf[i], &buf[j]);
}
}
};
// Generator to extend 64-bit seed values into longer sequences.
//
// The number of cycles is thus limited to 64-bits regardless of the engine, but this
// is still plenty for practical purposes.
const SplitMix64 = struct {
s: u64,
pub fn init(seed: u64) SplitMix64 {
return SplitMix64 { .s = seed };
}
pub fn next(self: &SplitMix64) u64 {
self.s +%= 0x9e3779b97f4a7c15;
var z = self.s;
z = (z ^ (z >> 30)) *% 0xbf58476d1ce4e5b9;
z = (z ^ (z >> 27)) *% 0x94d049bb133111eb;
return z ^ (z >> 31);
}
};
test "splitmix64 sequence" {
var r = SplitMix64.init(0xaeecf86f7878dd75);
const seq = []const u64 {
0x5dbd39db0178eb44,
0xa9900fb66b397da3,
0x5c1a28b1aeebcf5c,
0x64a963238f776912,
0xc6d4177b21d1c0ab,
0xb2cbdbdb5ea35394,
};
for (seq) |s| {
std.debug.assert(s == r.next());
}
}
// PCG32 - http://www.pcg-random.org/
//
// PRNG
pub const Pcg = struct {
const default_multiplier = 6364136223846793005;
random: Random,
s: u64,
i: u64,
pub fn init(init_s: u64) Pcg {
var pcg = Pcg {
.random = Random { .fillFn = fill },
.s = undefined,
.i = undefined,
};
pcg.seed(init_s);
return pcg;
}
fn next(self: &Pcg) u32 {
const l = self.s;
self.s = l *% default_multiplier +% (self.i | 1);
const xor_s = @truncate(u32, ((l >> 18) ^ l) >> 27);
const rot = u32(l >> 59);
return (xor_s >> u5(rot)) | (xor_s << u5((0 -% rot) & 31));
}
fn seed(self: &Pcg, init_s: u64) void {
// Pcg requires 128-bits of seed.
var gen = SplitMix64.init(init_s);
self.seedTwo(gen.next(), gen.next());
}
fn seedTwo(self: &Pcg, init_s: u64, init_i: u64) void {
self.s = 0;
self.i = (init_s << 1) | 1;
self.s = self.s *% default_multiplier +% self.i;
self.s +%= init_i;
self.s = self.s *% default_multiplier +% self.i;
}
fn fill(r: &Random, buf: []u8) void {
const self = @fieldParentPtr(Pcg, "random", r);
var i: usize = 0;
const aligned_len = buf.len - (buf.len & 7);
// Complete 4 byte segments.
while (i < aligned_len) : (i += 4) {
var n = self.next();
comptime var j: usize = 0;
inline while (j < 4) : (j += 1) {
buf[i + j] = @truncate(u8, n);
n >>= 8;
}
}
// Remaining. (cuts the stream)
if (i != buf.len) {
var n = self.next();
while (i < buf.len) : (i += 1) {
buf[i] = @truncate(u8, n);
n >>= 4;
}
}
}
};
test "pcg sequence" {
var r = Pcg.init(0);
const s0: u64 = 0x9394bf54ce5d79de;
const s1: u64 = 0x84e9c579ef59bbf7;
r.seedTwo(s0, s1);
const seq = []const u32 {
2881561918,
3063928540,
1199791034,
2487695858,
1479648952,
3247963454,
};
for (seq) |s| {
std.debug.assert(s == r.next());
}
}
// Xoroshiro128+ - http://xoroshiro.di.unimi.it/
//
// PRNG
pub const Xoroshiro128 = struct {
random: Random,
s: [2]u64,
pub fn init(init_s: u64) Xoroshiro128 {
var x = Xoroshiro128 {
.random = Random { .fillFn = fill },
.s = undefined,
};
x.seed(init_s);
return x;
}
fn next(self: &Xoroshiro128) u64 {
const s0 = self.s[0];
var s1 = self.s[1];
const r = s0 +% s1;
s1 ^= s0;
self.s[0] = math.rotl(u64, s0, u8(55)) ^ s1 ^ (s1 << 14);
self.s[1] = math.rotl(u64, s1, u8(36));
return r;
}
// Skip 2^64 places ahead in the sequence
fn jump(self: &Xoroshiro128) void {
var s0: u64 = 0;
var s1: u64 = 0;
const table = []const u64 {
0xbeac0467eba5facb,
0xd86b048b86aa9922
};
inline for (table) |entry| {
var b: usize = 0;
while (b < 64) : (b += 1) {
if ((entry & (u64(1) << u6(b))) != 0) {
s0 ^= self.s[0];
s1 ^= self.s[1];
}
_ = self.next();
}
}
self.s[0] = s0;
self.s[1] = s1;
}
fn seed(self: &Xoroshiro128, init_s: u64) void {
// Xoroshiro requires 128-bits of seed.
var gen = SplitMix64.init(init_s);
self.s[0] = gen.next();
self.s[1] = gen.next();
}
fn fill(r: &Random, buf: []u8) void {
const self = @fieldParentPtr(Xoroshiro128, "random", r);
var i: usize = 0;
const aligned_len = buf.len - (buf.len & 7);
// Complete 8 byte segments.
while (i < aligned_len) : (i += 8) {
var n = self.next();
comptime var j: usize = 0;
inline while (j < 8) : (j += 1) {
buf[i + j] = @truncate(u8, n);
n >>= 8;
}
}
// Remaining. (cuts the stream)
if (i != buf.len) {
var n = self.next();
while (i < buf.len) : (i += 1) {
buf[i] = @truncate(u8, n);
n >>= 8;
}
}
}
};
test "xoroshiro sequence" {
var r = Xoroshiro128.init(0);
r.s[0] = 0xaeecf86f7878dd75;
r.s[1] = 0x01cd153642e72622;
const seq1 = []const u64 {
0xb0ba0da5bb600397,
0x18a08afde614dccc,
0xa2635b956a31b929,
0xabe633c971efa045,
0x9ac19f9706ca3cac,
0xf62b426578c1e3fb,
};
for (seq1) |s| {
std.debug.assert(s == r.next());
}
r.jump();
const seq2 = []const u64 {
0x95344a13556d3e22,
0xb4fb32dafa4d00df,
0xb2011d9ccdcfe2dd,
0x05679a9b2119b908,
0xa860a1da7c9cd8a0,
0x658a96efe3f86550,
};
for (seq2) |s| {
std.debug.assert(s == r.next());
}
}
// ISAAC64 - http://www.burtleburtle.net/bob/rand/isaacafa.html
//
// CSPRNG
//
// Follows the general idea of the implementation from here with a few shortcuts.
// https://doc.rust-lang.org/rand/src/rand/prng/isaac64.rs.html
pub const Isaac64 = struct {
random: Random,
r: [256]u64,
m: [256]u64,
a: u64,
b: u64,
c: u64,
i: usize,
pub fn init(init_s: u64) Isaac64 {
var isaac = Isaac64 {
.random = Random { .fillFn = fill },
.r = undefined,
.m = undefined,
.a = undefined,
.b = undefined,
.c = undefined,
.i = undefined,
};
// seed == 0 => same result as the unseeded reference implementation
isaac.seed(init_s, 1);
return isaac;
}
fn step(self: &Isaac64, mix: u64, base: usize, comptime m1: usize, comptime m2: usize) void {
const x = self.m[base + m1];
self.a = mix +% self.m[base + m2];
const y = self.a +% self.b +% self.m[(x >> 3) % self.m.len];
self.m[base + m1] = y;
self.b = x +% self.m[(y >> 11) % self.m.len];
self.r[self.r.len - 1 - base - m1] = self.b;
}
fn refill(self: &Isaac64) void {
const midpoint = self.r.len / 2;
self.c +%= 1;
self.b +%= self.c;
{
var i: usize = 0;
while (i < midpoint) : (i += 4) {
self.step( ~(self.a ^ (self.a << 21)), i + 0, 0, midpoint);
self.step( self.a ^ (self.a >> 5) , i + 1, 0, midpoint);
self.step( self.a ^ (self.a << 12) , i + 2, 0, midpoint);
self.step( self.a ^ (self.a >> 33) , i + 3, 0, midpoint);
}
}
{
var i: usize = 0;
while (i < midpoint) : (i += 4) {
self.step( ~(self.a ^ (self.a << 21)), i + 0, midpoint, 0);
self.step( self.a ^ (self.a >> 5) , i + 1, midpoint, 0);
self.step( self.a ^ (self.a << 12) , i + 2, midpoint, 0);
self.step( self.a ^ (self.a >> 33) , i + 3, midpoint, 0);
}
}
self.i = 0;
}
fn next(self: &Isaac64) u64 {
if (self.i >= self.r.len) {
self.refill();
}
const value = self.r[self.i];
self.i += 1;
return value;
}
fn seed(self: &Isaac64, init_s: u64, comptime rounds: usize) void {
// We ignore the multi-pass requirement since we don't currently expose full access to
// seeding the self.m array completely.
mem.set(u64, self.m[0..], 0);
self.m[0] = init_s;
// prescrambled golden ratio constants
var a = []const u64 {
0x647c4677a2884b7c,
0xb9f8b322c73ac862,
0x8c0ea5053d4712a0,
0xb29b2e824a595524,
0x82f053db8355e0ce,
0x48fe4a0fa5a09315,
0xae985bf2cbfc89ed,
0x98f5704f6c44c0ab,
};
comptime var i: usize = 0;
inline while (i < rounds) : (i += 1) {
var j: usize = 0;
while (j < self.m.len) : (j += 8) {
comptime var x1: usize = 0;
inline while (x1 < 8) : (x1 += 1) {
a[x1] +%= self.m[j + x1];
}
a[0] -%= a[4]; a[5] ^= a[7] >> 9; a[7] +%= a[0];
a[1] -%= a[5]; a[6] ^= a[0] << 9; a[0] +%= a[1];
a[2] -%= a[6]; a[7] ^= a[1] >> 23; a[1] +%= a[2];
a[3] -%= a[7]; a[0] ^= a[2] << 15; a[2] +%= a[3];
a[4] -%= a[0]; a[1] ^= a[3] >> 14; a[3] +%= a[4];
a[5] -%= a[1]; a[2] ^= a[4] << 20; a[4] +%= a[5];
a[6] -%= a[2]; a[3] ^= a[5] >> 17; a[5] +%= a[6];
a[7] -%= a[3]; a[4] ^= a[6] << 14; a[6] +%= a[7];
comptime var x2: usize = 0;
inline while (x2 < 8) : (x2 += 1) {
self.m[j + x2] = a[x2];
}
}
}
mem.set(u64, self.r[0..], 0);
self.a = 0;
self.b = 0;
self.c = 0;
self.i = self.r.len; // trigger refill on first value
}
fn fill(r: &Random, buf: []u8) void {
const self = @fieldParentPtr(Isaac64, "random", r);
var i: usize = 0;
const aligned_len = buf.len - (buf.len & 7);
// Fill complete 64-byte segments
while (i < aligned_len) : (i += 8) {
var n = self.next();
comptime var j: usize = 0;
inline while (j < 8) : (j += 1) {
buf[i + j] = @truncate(u8, n);
n >>= 8;
}
}
// Fill trailing, ignoring excess (cut the stream).
if (i != buf.len) {
var n = self.next();
while (i < buf.len) : (i += 1) {
buf[i] = @truncate(u8, n);
n >>= 8;
}
}
}
};
test "isaac64 sequence" {
var r = Isaac64.init(0);
// from reference implementation
const seq = []const u64 {
0xf67dfba498e4937c,
0x84a5066a9204f380,
0xfee34bd5f5514dbb,
0x4d1664739b8f80d6,
0x8607459ab52a14aa,
0x0e78bc5a98529e49,
0xfe5332822ad13777,
0x556c27525e33d01a,
0x08643ca615f3149f,
0xd0771faf3cb04714,
0x30e86f68a37b008d,
0x3074ebc0488a3adf,
0x270645ea7a2790bc,
0x5601a0a8d3763c6a,
0x2f83071f53f325dd,
0xb9090f3d42d2d2ea,
};
for (seq) |s| {
std.debug.assert(s == r.next());
}
}
// Actual Random helper function tests, pcg engine is assumed correct.
test "Random float" {
var prng = DefaultPrng.init(0);
var i: usize = 0;
while (i < 1000) : (i += 1) {
const val1 = prng.random.float(f32);
std.debug.assert(val1 >= 0.0);
std.debug.assert(val1 < 1.0);
const val2 = prng.random.float(f64);
std.debug.assert(val2 >= 0.0);
std.debug.assert(val2 < 1.0);
}
}
test "Random scalar" {
var prng = DefaultPrng.init(0);
const s = prng .random.scalar(u64);
}
test "Random bytes" {
var prng = DefaultPrng.init(0);
var buf: [2048]u8 = undefined;
prng.random.bytes(buf[0..]);
}
test "Random shuffle" {
var prng = DefaultPrng.init(0);
var seq = []const u8 { 0, 1, 2, 3, 4 };
var seen = []bool {false} ** 5;
var i: usize = 0;
while (i < 1000) : (i += 1) {
prng.random.shuffle(u8, seq[0..]);
seen[seq[0]] = true;
std.debug.assert(sumArray(seq[0..]) == 10);
}
// we should see every entry at the head at least once
for (seen) |e| {
std.debug.assert(e == true);
}
}
fn sumArray(s: []const u8) u32 {
var r: u32 = 0;
for (s) |e| r += e;
return r;
}
test "Random range" {
var prng = DefaultPrng.init(0);
testRange(&prng.random, -4, 3);
testRange(&prng.random, -4, -1);
testRange(&prng.random, 10, 14);
}
fn testRange(r: &Random, start: i32, end: i32) void {
const count = usize(end - start);
var values_buffer = []bool{false} ** 20;
const values = values_buffer[0..count];
var i: usize = 0;
while (i < count) {
const value = r.range(i32, start, end);
const index = usize(value - start);
if (!values[index]) {
i += 1;
values[index] = true;
}
}
}