For the original math/rand, we ported Plan 9's random number
generator, which was a refinement by Ken Thompson of an algorithm
by Don Mitchell and Jim Reeds, which Mitchell in turn recalls as
having been derived from an algorithm by Marsaglia. At its core,
it is an additive lagged Fibonacci generator (ALFG).
Whatever the details of the history, this generator is nowhere
near the current state of the art for simple, pseudo-random
generators.
This CL adds an implementation of Melissa O'Neill's PCG, specifically
the variant PCG-DXSM, which she defined after writing the PCG paper
and which is now the default in Numpy. The update is slightly slower
(a few multiplies and adds, instead of a few adds), but the state
is dramatically smaller (2 words instead of 607). The statistical
output properties are better too.
A followup CL will delete the old generator.
PCG is the only change here, so no benchmarks should be affected.
Including them anyway as further evidence for caution.
goos: linux
goarch: amd64
pkg: math/rand/v2
cpu: AMD Ryzen 9 7950X 16-Core Processor
│
8993506f2f.amd64 │
01ff938549.amd64 │
│ sec/op │ sec/op vs base │
SourceUint64-32 1.325n ± 1% 1.352n ± 1% +2.00% (p=0.000 n=20)
GlobalInt64-32 2.240n ± 1% 2.083n ± 0% -7.03% (p=0.000 n=20)
GlobalInt64Parallel-32 0.1041n ± 1% 0.1035n ± 1% ~ (p=0.064 n=20)
GlobalUint64-32 2.072n ± 3% 2.038n ± 1% ~ (p=0.089 n=20)
GlobalUint64Parallel-32 0.1008n ± 1% 0.1006n ± 1% ~ (p=0.804 n=20)
Int64-32 1.716n ± 1% 1.687n ± 2% ~ (p=0.045 n=20)
Uint64-32 1.665n ± 1% 1.674n ± 2% ~ (p=0.878 n=20)
GlobalIntN1000-32 3.335n ± 1% 3.135n ± 1% -6.00% (p=0.000 n=20)
IntN1000-32 2.484n ± 1% 2.478n ± 1% ~ (p=0.085 n=20)
Int64N1000-32 2.502n ± 2% 2.455n ± 1% -1.88% (p=0.002 n=20)
Int64N1e8-32 2.484n ± 2% 2.467n ± 2% ~ (p=0.048 n=20)
Int64N1e9-32 2.502n ± 0% 2.454n ± 1% -1.92% (p=0.000 n=20)
Int64N2e9-32 2.502n ± 0% 2.482n ± 1% -0.76% (p=0.000 n=20)
Int64N1e18-32 3.201n ± 1% 3.349n ± 2% +4.62% (p=0.000 n=20)
Int64N2e18-32 3.504n ± 1% 3.537n ± 1% ~ (p=0.185 n=20)
Int64N4e18-32 4.873n ± 1% 4.917n ± 0% +0.90% (p=0.000 n=20)
Int32N1000-32 2.639n ± 1% 2.386n ± 1% -9.57% (p=0.000 n=20)
Int32N1e8-32 2.686n ± 2% 2.366n ± 1% -11.91% (p=0.000 n=20)
Int32N1e9-32 2.636n ± 1% 2.355n ± 2% -10.70% (p=0.000 n=20)
Int32N2e9-32 2.660n ± 1% 2.371n ± 1% -10.88% (p=0.000 n=20)
Float32-32 2.261n ± 1% 2.245n ± 2% ~ (p=0.752 n=20)
Float64-32 2.280n ± 1% 2.235n ± 1% -1.97% (p=0.007 n=20)
ExpFloat64-32 3.891n ± 1% 3.813n ± 3% ~ (p=0.087 n=20)
NormFloat64-32 3.711n ± 1% 3.652n ± 2% ~ (p=0.021 n=20)
Perm3-32 32.60n ± 2% 33.12n ± 3% ~ (p=0.107 n=20)
Perm30-32 204.2n ± 0% 205.1n ± 1% ~ (p=0.358 n=20)
Perm30ViaShuffle-32 121.7n ± 2% 110.8n ± 1% -8.96% (p=0.000 n=20)
ShuffleOverhead-32 106.2n ± 2% 113.0n ± 1% +6.36% (p=0.000 n=20)
Concurrent-32 2.190n ± 5% 2.100n ± 0% -4.13% (p=0.001 n=20)
PCG_DXSM-32 1.490n ± 0%
goos: darwin
goarch: arm64
pkg: math/rand/v2
cpu: Apple M1
│
8993506f2f.arm64 │
01ff938549.arm64 │
│ sec/op │ sec/op vs base │
SourceUint64-8 2.271n ± 0% 2.258n ± 1% ~ (p=0.167 n=20)
GlobalInt64-8 2.161n ± 1% 2.167n ± 0% ~ (p=0.693 n=20)
GlobalInt64Parallel-8 0.4303n ± 0% 0.4310n ± 0% ~ (p=0.051 n=20)
GlobalUint64-8 2.164n ± 1% 2.182n ± 1% ~ (p=0.042 n=20)
GlobalUint64Parallel-8 0.4287n ± 0% 0.4297n ± 0% ~ (p=0.082 n=20)
Int64-8 2.478n ± 1% 2.472n ± 1% ~ (p=0.151 n=20)
Uint64-8 2.460n ± 1% 2.449n ± 1% ~ (p=0.013 n=20)
GlobalIntN1000-8 2.814n ± 2% 2.814n ± 2% ~ (p=0.821 n=20)
IntN1000-8 3.003n ± 2% 2.998n ± 2% ~ (p=0.024 n=20)
Int64N1000-8 2.954n ± 0% 2.949n ± 2% ~ (p=0.192 n=20)
Int64N1e8-8 2.956n ± 0% 2.953n ± 2% ~ (p=0.109 n=20)
Int64N1e9-8 3.325n ± 0% 2.950n ± 0% -11.26% (p=0.000 n=20)
Int64N2e9-8 2.956n ± 2% 2.946n ± 2% ~ (p=0.027 n=20)
Int64N1e18-8 3.780n ± 1% 3.779n ± 1% ~ (p=0.815 n=20)
Int64N2e18-8 4.385n ± 0% 4.370n ± 1% ~ (p=0.402 n=20)
Int64N4e18-8 6.527n ± 0% 6.544n ± 1% ~ (p=0.140 n=20)
Int32N1000-8 2.964n ± 1% 2.950n ± 0% -0.47% (p=0.002 n=20)
Int32N1e8-8 2.964n ± 1% 2.950n ± 2% ~ (p=0.013 n=20)
Int32N1e9-8 2.963n ± 2% 2.951n ± 2% ~ (p=0.062 n=20)
Int32N2e9-8 2.961n ± 2% 2.950n ± 2% -0.37% (p=0.002 n=20)
Float32-8 3.442n ± 0% 3.441n ± 0% ~ (p=0.211 n=20)
Float64-8 3.442n ± 0% 3.442n ± 0% ~ (p=0.067 n=20)
ExpFloat64-8 4.472n ± 0% 4.481n ± 0% +0.20% (p=0.000 n=20)
NormFloat64-8 4.734n ± 0% 4.725n ± 0% -0.19% (p=0.003 n=20)
Perm3-8 26.55n ± 0% 26.55n ± 0% ~ (p=0.833 n=20)
Perm30-8 181.9n ± 0% 181.9n ± 0% -0.03% (p=0.004 n=20)
Perm30ViaShuffle-8 143.1n ± 0% 142.9n ± 0% ~ (p=0.204 n=20)
ShuffleOverhead-8 120.6n ± 1% 120.8n ± 2% ~ (p=0.102 n=20)
Concurrent-8 2.357n ± 2% 2.421n ± 6% ~ (p=0.016 n=20)
PCG_DXSM-8 2.531n ± 0%
goos: linux
goarch: 386
pkg: math/rand/v2
cpu: AMD Ryzen 9 7950X 16-Core Processor
│
8993506f2f.386 │
01ff938549.386 │
│ sec/op │ sec/op vs base │
SourceUint64-32 2.102n ± 2% 2.069n ± 0% ~ (p=0.021 n=20)
GlobalInt64-32 3.542n ± 2% 3.456n ± 1% -2.44% (p=0.001 n=20)
GlobalInt64Parallel-32 0.3202n ± 0% 0.3252n ± 0% +1.56% (p=0.000 n=20)
GlobalUint64-32 3.507n ± 1% 3.573n ± 1% +1.87% (p=0.000 n=20)
GlobalUint64Parallel-32 0.3170n ± 1% 0.3159n ± 0% ~ (p=0.167 n=20)
Int64-32 2.516n ± 1% 2.562n ± 2% ~ (p=0.016 n=20)
Uint64-32 2.544n ± 1% 2.592n ± 0% +1.85% (p=0.000 n=20)
GlobalIntN1000-32 6.237n ± 1% 6.266n ± 2% ~ (p=0.268 n=20)
IntN1000-32 4.670n ± 2% 4.724n ± 2% ~ (p=0.644 n=20)
Int64N1000-32 5.412n ± 1% 5.490n ± 2% ~ (p=0.159 n=20)
Int64N1e8-32 5.414n ± 2% 5.513n ± 2% ~ (p=0.129 n=20)
Int64N1e9-32 5.473n ± 1% 5.476n ± 1% ~ (p=0.723 n=20)
Int64N2e9-32 5.487n ± 1% 5.501n ± 2% ~ (p=0.481 n=20)
Int64N1e18-32 8.901n ± 2% 9.043n ± 2% ~ (p=0.330 n=20)
Int64N2e18-32 9.521n ± 1% 9.601n ± 2% ~ (p=0.703 n=20)
Int64N4e18-32 11.92n ± 1% 12.00n ± 1% ~ (p=0.489 n=20)
Int32N1000-32 4.785n ± 1% 4.829n ± 2% ~ (p=0.402 n=20)
Int32N1e8-32 4.748n ± 1% 4.825n ± 2% ~ (p=0.218 n=20)
Int32N1e9-32 4.810n ± 1% 4.830n ± 2% ~ (p=0.794 n=20)
Int32N2e9-32 4.812n ± 1% 4.750n ± 2% ~ (p=0.057 n=20)
Float32-32 10.48n ± 4% 10.89n ± 4% ~ (p=0.162 n=20)
Float64-32 19.79n ± 3% 19.60n ± 4% ~ (p=0.668 n=20)
ExpFloat64-32 12.91n ± 3% 12.96n ± 3% ~ (p=1.000 n=20)
NormFloat64-32 7.462n ± 1% 7.516n ± 1% ~ (p=0.051 n=20)
Perm3-32 35.98n ± 2% 36.78n ± 2% ~ (p=0.033 n=20)
Perm30-32 241.5n ± 1% 238.9n ± 2% ~ (p=0.126 n=20)
Perm30ViaShuffle-32 187.3n ± 2% 189.7n ± 2% ~ (p=0.387 n=20)
ShuffleOverhead-32 160.2n ± 1% 159.8n ± 1% ~ (p=0.256 n=20)
Concurrent-32 3.308n ± 3% 3.286n ± 1% ~ (p=0.038 n=20)
PCG_DXSM-32 7.613n ± 1%
For #61716.
Change-Id: Icb274ca1f782504d658305a40159b4ae6a2f3f1d
Reviewed-on: https://go-review.googlesource.com/c/go/+/502505
Auto-Submit: Russ Cox <rsc@golang.org>
Reviewed-by: Dmitri Shuralyov <dmitshur@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Rob Pike <r@golang.org>
pkg math/rand/v2, func IntN(int) int #61716
pkg math/rand/v2, func N[$0 intType]($0) $0 #61716
pkg math/rand/v2, func New(Source) *Rand #61716
+pkg math/rand/v2, func NewPCG(uint64, uint64) *PCG #61716
pkg math/rand/v2, func NewSource(int64) Source #61716
pkg math/rand/v2, func NewZipf(*Rand, float64, float64, uint64) *Zipf #61716
pkg math/rand/v2, func NormFloat64() float64 #61716
pkg math/rand/v2, func Uint64() uint64 #61716
pkg math/rand/v2, func Uint64N(uint64) uint64 #61716
pkg math/rand/v2, func UintN(uint) uint #61716
+pkg math/rand/v2, method (*PCG) MarshalBinary() ([]uint8, error) #61716
+pkg math/rand/v2, method (*PCG) Seed(uint64, uint64) #61716
+pkg math/rand/v2, method (*PCG) Uint64() uint64 #61716
+pkg math/rand/v2, method (*PCG) UnmarshalBinary([]uint8) error #61716
pkg math/rand/v2, method (*Rand) ExpFloat64() float64 #61716
pkg math/rand/v2, method (*Rand) Float32() float32 #61716
pkg math/rand/v2, method (*Rand) Float64() float64 #61716
pkg math/rand/v2, method (*Rand) Uint64N(uint64) uint64 #61716
pkg math/rand/v2, method (*Rand) UintN(uint) uint #61716
pkg math/rand/v2, method (*Zipf) Uint64() uint64 #61716
+pkg math/rand/v2, type PCG struct #61716
pkg math/rand/v2, type Rand struct #61716
pkg math/rand/v2, type Source interface { Uint64 } #61716
pkg math/rand/v2, type Source interface, Uint64() uint64 #61716
--- /dev/null
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package rand
+
+import (
+ "errors"
+ "math/bits"
+)
+
+// https://numpy.org/devdocs/reference/random/upgrading-pcg64.html
+// https://github.com/imneme/pcg-cpp/commit/871d0494ee9c9a7b7c43f753e3d8ca47c26f8005
+
+// A PCG is a PCG generator with 128 bits of internal state.
+// A zero PCG is equivalent to NewPCG(0, 0).
+type PCG struct {
+ hi uint64
+ lo uint64
+}
+
+// NewPCG returns a new PCG seeded with the given values.
+func NewPCG(seed1, seed2 uint64) *PCG {
+ return &PCG{seed1, seed2}
+}
+
+// Seed resets the PCG to behave the same way as NewPCG(seed1, seed2).
+func (p *PCG) Seed(seed1, seed2 uint64) {
+ p.hi = seed1
+ p.lo = seed2
+}
+
+// binary.bigEndian.Uint64, copied to avoid dependency
+func beUint64(b []byte) uint64 {
+ _ = b[7] // bounds check hint to compiler; see golang.org/issue/14808
+ return uint64(b[7]) | uint64(b[6])<<8 | uint64(b[5])<<16 | uint64(b[4])<<24 |
+ uint64(b[3])<<32 | uint64(b[2])<<40 | uint64(b[1])<<48 | uint64(b[0])<<56
+}
+
+// binary.bigEndian.PutUint64, copied to avoid dependency
+func bePutUint64(b []byte, v uint64) {
+ _ = b[7] // early bounds check to guarantee safety of writes below
+ b[0] = byte(v >> 56)
+ b[1] = byte(v >> 48)
+ b[2] = byte(v >> 40)
+ b[3] = byte(v >> 32)
+ b[4] = byte(v >> 24)
+ b[5] = byte(v >> 16)
+ b[6] = byte(v >> 8)
+ b[7] = byte(v)
+}
+
+// MarshalBinary implements the encoding.BinaryMarshaler interface.
+func (p *PCG) MarshalBinary() ([]byte, error) {
+ b := make([]byte, 20)
+ copy(b, "pcg:")
+ bePutUint64(b[4:], p.hi)
+ bePutUint64(b[4+8:], p.lo)
+ return b, nil
+}
+
+var errUnmarshalPCG = errors.New("invalid PCG encoding")
+
+// UnmarshalBinary implements the encoding.BinaryUnmarshaler interface.
+func (p *PCG) UnmarshalBinary(data []byte) error {
+ if len(data) != 20 || string(data[:4]) != "pcg:" {
+ return errUnmarshalPCG
+ }
+ p.hi = beUint64(data[4:])
+ p.lo = beUint64(data[4+8:])
+ return nil
+}
+
+func (p *PCG) next() (hi, lo uint64) {
+ // https://github.com/imneme/pcg-cpp/blob/428802d1a5/include/pcg_random.hpp#L161
+ //
+ // Numpy's PCG multiplies by the 64-bit value cheapMul
+ // instead of the 128-bit value used here and in the official PCG code.
+ // This does not seem worthwhile, at least for Go: not having any high
+ // bits in the multiplier reduces the effect of low bits on the highest bits,
+ // and it only saves 1 multiply out of 3.
+ // (On 32-bit systems, it saves 1 out of 6, since Mul64 is doing 4.)
+ const (
+ mulHi = 2549297995355413924
+ mulLo = 4865540595714422341
+ incHi = 6364136223846793005
+ incLo = 1442695040888963407
+ )
+
+ // state = state * mul + inc
+ hi, lo = bits.Mul64(p.lo, mulLo)
+ hi += p.hi*mulLo + p.lo*mulHi
+ lo, c := bits.Add64(lo, incLo, 0)
+ hi, _ = bits.Add64(hi, incHi, c)
+ p.lo = lo
+ p.hi = hi
+ return hi, lo
+}
+
+// Uint64 return a uniformly-distributed random uint64 value.
+func (p *PCG) Uint64() uint64 {
+ hi, lo := p.next()
+
+ // XSL-RR would be
+ // hi, lo := p.next()
+ // return bits.RotateLeft64(lo^hi, -int(hi>>58))
+ // but Numpy uses DXSM and O'Neill suggests doing the same.
+ // See https://github.com/golang/go/issues/21835#issuecomment-739065688
+ // and following comments.
+
+ // DXSM "double xorshift multiply"
+ // https://github.com/imneme/pcg-cpp/blob/428802d1a5/include/pcg_random.hpp#L1015
+
+ // https://github.com/imneme/pcg-cpp/blob/428802d1a5/include/pcg_random.hpp#L176
+ const cheapMul = 0xda942042e4dd58b5
+ hi ^= hi >> 32
+ hi *= cheapMul
+ hi ^= hi >> 48
+ hi *= (lo | 1)
+ return hi
+}
--- /dev/null
+// Copyright 2023 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package rand_test
+
+import (
+ . "math/rand/v2"
+ "testing"
+)
+
+func BenchmarkPCG_DXSM(b *testing.B) {
+ var p PCG
+ var t uint64
+ for n := b.N; n > 0; n-- {
+ t += p.Uint64()
+ }
+ Sink = t
+}
+
+func TestPCGMarshal(t *testing.T) {
+ var p PCG
+ const (
+ seed1 = 0x123456789abcdef0
+ seed2 = 0xfedcba9876543210
+ want = "pcg:\x12\x34\x56\x78\x9a\xbc\xde\xf0\xfe\xdc\xba\x98\x76\x54\x32\x10"
+ )
+ p.Seed(seed1, seed2)
+ data, err := p.MarshalBinary()
+ if string(data) != want || err != nil {
+ t.Errorf("MarshalBinary() = %q, %v, want %q, nil", data, err, want)
+ }
+
+ q := PCG{}
+ if err := q.UnmarshalBinary([]byte(want)); err != nil {
+ t.Fatalf("UnmarshalBinary(): %v", err)
+ }
+ if q != p {
+ t.Fatalf("after round trip, q = %#x, but p = %#x", q, p)
+ }
+
+ qu := q.Uint64()
+ pu := p.Uint64()
+ if qu != pu {
+ t.Errorf("after round trip, q.Uint64() = %#x, but p.Uint64() = %#x", qu, pu)
+ }
+}
+
+func TestPCG(t *testing.T) {
+ p := NewPCG(1, 2)
+ want := []uint64{
+ 0xc4f5a58656eef510,
+ 0x9dcec3ad077dec6c,
+ 0xc8d04605312f8088,
+ 0xcbedc0dcb63ac19a,
+ 0x3bf98798cae97950,
+ 0xa8c6d7f8d485abc,
+ 0x7ffa3780429cd279,
+ 0x730ad2626b1c2f8e,
+ 0x21ff2330f4a0ad99,
+ 0x2f0901a1947094b0,
+ 0xa9735a3cfbe36cef,
+ 0x71ddb0a01a12c84a,
+ 0xf0e53e77a78453bb,
+ 0x1f173e9663be1e9d,
+ 0x657651da3ac4115e,
+ 0xc8987376b65a157b,
+ 0xbb17008f5fca28e7,
+ 0x8232bd645f29ed22,
+ 0x12be8f07ad14c539,
+ 0x54908a48e8e4736e,
+ }
+
+ for i, x := range want {
+ if u := p.Uint64(); u != x {
+ t.Errorf("PCG #%d = %#x, want %#x", i, u, x)
+ }
+ }
+}