math.big: add a new greatest-common-divisor-algo for big.Integer, also add a benchmark for it (#12261)

pull/12301/head
Andreas Schoeller 2021-10-26 10:10:13 +02:00 committed by GitHub
parent f62b2dcfa7
commit f14dabc6bd
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 633 additions and 0 deletions

1
.gitignore vendored
View File

@ -20,6 +20,7 @@ vdbg.exe
*.dll *.dll
*.lib *.lib
*.bak *.bak
*.dylib
a.out a.out
.noprefix.vrepl_temp .noprefix.vrepl_temp

View File

@ -761,3 +761,71 @@ pub fn (a Integer) isqrt() Integer {
} }
return result return result
} }
[inline]
fn bi_min(a Integer, b Integer) Integer {
return if a < b { a } else { b }
}
[inline]
fn bi_max(a Integer, b Integer) Integer {
return if a > b { a } else { b }
}
[direct_array_access]
fn (bi Integer) msb() u32 {
for idx := 0; idx < bi.digits.len; idx += 1 {
word := bi.digits[idx]
if word > 0 {
return u32((idx * 32) + bits.trailing_zeros_32(word))
}
}
return u32(32)
}
// Greatest-Common-Divisor https://en.wikipedia.org/wiki/Binary_GCD_algorithm
// The code below follows the 2013-christmas-special by D. Lemire & R. Corderoy
// https://en.algorithmica.org/hpc/analyzing-performance/gcd/
//
// discussion & further info https://lemire.me/blog/2013/12/26/fastest-way-to-compute-the-greatest-common-divisor/
pub fn (x Integer) gcd_binary(y Integer) Integer {
// Since standard-euclid-gcd is much faster on smaller sizes 4-8-Byte.
// In such a case, one could delegate back to big.Integer.gcd()
// Uncomment below and a all long-long goes to euclid-gcd.
//
// if x.digits.len + y.digits.len <= 4 {
// return x.gcd( y )
// }
if x.signum == 0 {
return y.abs()
}
if y.signum == 0 {
return x.abs()
}
if x.signum < 0 {
return x.neg().gcd(y)
}
if y.signum < 0 {
return x.gcd(y.neg())
}
mut a := x
mut b := y
mut az := a.msb()
bz := b.msb()
shift := util.umin(az, bz)
b = b.rshift(bz)
for a.signum != 0 {
a = a.rshift(az)
diff := b - a
az = diff.msb()
b = bi_min(a, b)
a = diff.abs()
}
return b.lshift(shift)
}

View File

@ -0,0 +1,339 @@
module main
// NB: this benchmark is preferable to be compiled with: `v -prod -cg -gc boehm bench_euclid.v`
import math.big
import benchmark
import os
import v.tests.bench.math_big_gcd.prime {
DataI,
PrimeCfg,
PrimeSet,
}
interface TestDataI {
r big.Integer
aa big.Integer
bb big.Integer
}
type GCDSet = PrimeSet
type Clocks = map[string]benchmark.Benchmark
const (
empty_set = GCDSet{'1', '1', '1'}
with_dots = false
)
fn main() {
fp := os.join_path(@VROOT, prime.toml_path)
if !prime_file_exists(fp) {
panic('expected file |$fp| - not found.')
}
mut clocks := Clocks(map[string]benchmark.Benchmark{})
for algo in [
'euclid',
'gcd_binary',
//'u32binary',
//'u64binary'
] {
clocks[algo] = benchmark.new_benchmark()
}
// any test-prime-set needs to pass this predicate
// before used in the benchmark.
//
predicate_fn := fn (ps PrimeSet) bool {
cast_bi := bi_from_decimal_string
r := cast_bi(ps.r)
aa := r * cast_bi(ps.a)
bb := r * cast_bi(ps.b)
gcd := aa.gcd(bb)
return if [
gcd != big.one_int,
gcd == bb.gcd(aa),
gcd == r,
aa != bb,
(aa % gcd) == big.zero_int,
(bb % gcd) == big.zero_int,
].all(it == true)
{ true } else { false }
}
cfgs := [
// root-prime a-prime b-prime
['s.3', 'xs.3', 's'],
['s.10', 's.10', 'm.all'],
['m.9', 's.10', 'xs'],
['l.40', 'xs.10', 's.5'],
['ml.20', 'm', 's.15'],
['xl', 's.10', 'l.15'],
['xxl', 'l.10', 'xl'],
['l.30', 'm.10', 'xxl'],
['xxl', 'xl', 'm.15'],
['mega', 'xl.6', 's'],
['xxl', 'xxxl.10', 'm.30'],
['mega', 'xxxl', 'mega'],
['giga.5', 'mega', 'giga'],
['crazy', 'mega', 'giga'],
['s', 'biggest', 'crazy'],
['biggest', 'crazy', 'giga'],
].map(PrimeCfg{it[0], it[1], it[2]})
println('\n$cfgs.len x Tests')
for i, prime_cfg in cfgs {
println('\n#-${i + 1}(Stack) "$prime_cfg"')
bench_euclid_vs_binary(prime_cfg, false, predicate_fn, mut clocks)
// just-to-be-sure, but makes no difference in this test.
// println('\n#-${i + 1}(Heap) "$prime_cfg"')
// bench_euclid_vs_binary(prime_cfg, true, predicate_fn, mut clocks)
}
println('')
for _, mut clock in clocks {
clock.stop()
}
println(clocks['euclid'].total_message('both algorithms '))
msg := [
'Seems to me as if euclid in big.Integer.gcd() performs better on ',
'very-small-integers up to 8-byte/u64. The tests #-1..5 show this.',
'The gcd_binary-algo seems to be perform better, the larger the numbers/buffers get.',
'On my machine, i see consistent gains between ~10-30-percent with :',
'\n',
'v -prod -cg -gc boehm bench_euclid.v',
'\n',
'This test covers multiplied primes up-to a length of 300-char-digits in ',
'a decimal-string. This equals (188-byte) == 47 x u32-values.',
'edit/change primes in $prime.toml_path',
'Improvements & critique are welcome : \n',
'https://lemire.me/blog/2013/12/26/fastest-way-to-compute-the-greatest-common-divisor/',
].join('\n')
println(msg)
}
fn run_benchmark(data []DataI, heap bool, mut clocks Clocks) bool {
mut testdata := []TestDataI{}
for elem in data {
// if elem is StackData || elem is HeapData{
// testdata << elem
// }
// TODO: this reads strange
if elem is StackData {
testdata << elem
}
if elem is HeapData {
testdata << elem
}
}
// some statistics
//
mut tmp := []int{cap: testdata.len * 3}
for set in testdata {
for prime in [set.r, set.aa, set.bb] {
bi, _ := prime.bytes()
tmp << bi_buffer_len(bi)
}
}
tmp.sort()
min_byte := tmp.first() * 4
max_byte := tmp.last() * 4
mut buffer_space := 0
for tmp.len != 0 {
buffer_space += tmp.pop()
}
// trying to balance prime-size and item-count
// minimum rounds is 100-times
//
mut rounds := 2000 / ((3 * buffer_space) / testdata.len)
rounds = if rounds < 100 { 100 } else { rounds }
ratio := (buffer_space * 4) / (testdata.len * 3)
msg := [
'avg-$ratio-byte/Prime, $min_byte-byte < Prime < $max_byte-byte \n',
'~${buffer_space * 4 / 1024}-Kb-',
if heap { 'Heap' } else { 'Stack' },
'-space for $testdata.len-items x ',
'$rounds-rounds',
].join('')
println(msg)
mut cycles := 0
for algo, mut clock in clocks {
cycles = 0
clock.step()
for cycles < rounds {
for set in testdata {
match algo {
'euclid' {
if set.r != set.aa.gcd(set.bb) {
eprintln('$algo failed ?')
clock.fail()
break
}
}
'gcd_binary' {
if set.r != set.aa.gcd_binary(set.bb) {
eprintln('$algo failed ?')
clock.fail()
break
}
}
else {
eprintln('unknown algo was "$algo"')
continue
}
}
} // eo-for over testdata
if with_dots {
print('.')
}
cycles += 1
} // eo-for cycles
if with_dots {
println('')
}
clock.ok()
clock.measure(algo)
} // eo-for-loop over algo, clock
return true
}
fn bench_euclid_vs_binary(test_config PrimeCfg, heap bool, predicate_fn fn (ps PrimeSet) bool, mut clocks Clocks) bool {
testprimes := prime.random_set(test_config) or { panic(err) }
// validate the test-data
//
gcd_primes := testprimes.map(prepare_and_test_gcd(it, predicate_fn)).filter(it != empty_set)
// just to make sure all generated primes are sane
//
assert gcd_primes.len == testprimes.len
// casting the decimal-strings into big.Integers
// here to avoid measuring string-parsing-cycles
// during later testing.
//
mut casted_sets := if heap { gcd_primes.map(unsafe {
DataI(&PrimeSet(&it)).cast<HeapData>()
}) } else { gcd_primes.map(unsafe {
DataI(&PrimeSet(&it)).cast<StackData>()
}) }
// ready use the primes in the benchmark
//
return run_benchmark(casted_sets, heap, mut clocks)
}
fn prepare_and_test_gcd(primeset PrimeSet, test fn (ps PrimeSet) bool) GCDSet {
if !primeset.predicate(test) {
eprintln('? Corrupt Testdata was ?')
dump(primeset)
return empty_set // {'1', '1', '1'}
}
cast_bi := bi_from_decimal_string
r := cast_bi(primeset.r)
aa := cast_bi(primeset.a) * r
bb := cast_bi(primeset.b) * r
gcd := aa.gcd(bb)
return GCDSet{'$gcd', '$aa', '$bb'}
}
fn prime_file_exists(path string) bool {
return os.is_readable(path)
}
// bi_from_decimal_string converts a string-of-decimals into
// a math.big.Integer using the big.Integers 'from_radix-fn'
//
pub fn bi_from_decimal_string(s string) big.Integer {
return big.integer_from_radix(s, u32(10)) or {
msg := [
'Cannot convert prime from decimal-string.',
'prime was : "$s"\n',
].join('\n')
panic(msg)
}
}
// need the bi.digits.len - during test only - to calculate
// the size of big.Integers-buffer
//
fn bi_buffer_len(input []byte) int {
if input.len == 0 {
return 0
}
// pad input
mut padded_input := []byte{len: ((input.len + 3) & ~0x3) - input.len, cap: (input.len + 3) & ~0x3, init: 0x0}
padded_input << input
mut digits := []u32{len: padded_input.len / 4}
// combine every 4 bytes into a u32 and insert into n.digits
for i := 0; i < padded_input.len; i += 4 {
x3 := u32(padded_input[i])
x2 := u32(padded_input[i + 1])
x1 := u32(padded_input[i + 2])
x0 := u32(padded_input[i + 3])
val := (x3 << 24) | (x2 << 16) | (x1 << 8) | x0
digits[(padded_input.len - i) / 4 - 1] = val
}
return digits.len
}
[heap]
pub struct HeapData {
pub mut:
r big.Integer
aa big.Integer
bb big.Integer
}
pub fn (hd HeapData) to_primeset() PrimeSet {
return PrimeSet{
r: '$hd.r'
a: '$hd.aa'
b: '$hd.bb'
}
}
pub fn (hd HeapData) from_primeset(p PrimeSet) DataI {
return DataI(HeapData{
r: bi_from_decimal_string(p.r)
aa: bi_from_decimal_string(p.a)
bb: bi_from_decimal_string(p.b)
})
}
pub struct StackData {
pub mut:
r big.Integer
aa big.Integer
bb big.Integer
}
pub fn (sd StackData) to_primeset() PrimeSet {
return PrimeSet{
r: '$sd.r'
a: '$sd.aa'
b: '$sd.bb'
}
}
pub fn (sd StackData) from_primeset(p PrimeSet) DataI {
return DataI(StackData{
r: bi_from_decimal_string(p.r)
aa: bi_from_decimal_string(p.a)
bb: bi_from_decimal_string(p.b)
})
}

View File

@ -0,0 +1,177 @@
module prime
import rand
import toml
import os
pub const toml_path = 'vlib/v/tests/bench/math_big_gcd/primes.toml'
pub interface DataI {
to_primeset() PrimeSet
from_primeset(PrimeSet) DataI
}
pub fn (di DataI) cast<T>() DataI {
return T{}.from_primeset(di.to_primeset())
}
pub type PrimeCfg = PrimeSet
pub fn (pc PrimeCfg) short() string {
return "r: '$pc.r' a: '$pc.a' b: '$pc.b'"
}
[heap]
pub struct PrimeSet {
pub mut:
r string [required]
a string [required]
b string [required]
}
pub fn (p PrimeSet) to_primeset() PrimeSet {
return p
}
pub fn (p PrimeSet) from_primeset(ps PrimeSet) DataI {
return ps
}
pub fn (p PrimeSet) predicate(pred fn (data PrimeSet) bool) bool {
return pred(p)
}
pub fn (p PrimeSet) key() string {
return [p.r, p.a, p.b].join('.')
}
pub fn (p PrimeSet) str() string {
return [p.r, p.a, p.b].join(' ')
}
fn extract_count(s string) int {
digits := '0123456789'.split('')
if (s == '') || !s.split('').any(it in digits) {
return 0
}
ds := s.split('').filter(it in digits)
return ds.join('').int()
}
// sizes lists the available names for prime-number sections.
pub fn sizes() []string {
primes := read_toml_file()
return primes.keys()
}
// usage returns section-names and the count of available primes
//
pub fn usage() string {
primes := read_toml_file()
return sizes().map('$it[..${primes[it].len}]').join('\n\t')
}
// reads the Map[string] []string from disk
// and returns the parsed content
fn read_toml_file() map[string][]string {
fp := os.join_path(@VROOT, prime.toml_path)
tm_doc := toml.parse_file(fp) or {
err_msg := 'expected $fp'
eprintln(err_msg)
panic(err)
}
// TODO what happens if this goes wrong ?
tm_primes := tm_doc.value('primes') as map[string]toml.Any
msg := 'expected a map[string][]string in TOML-data ? corrupt ?'
mut p := map[string][]string{}
for k in tm_primes.keys() {
p[k] = []string{}
arr := tm_primes[k] or { panic(msg) }
for _, elem in arr.array() {
p[k] << elem as string
}
}
return p
}
pub fn random_list(cfg []string) []string {
primes := read_toml_file()
mut p_list := []string{}
match cfg.len {
1 { // prime-size e.g. 'xs' given
if cfg[0] !in primes {
return p_list
} else {
return primes[cfg[0]]
}
}
2 { // prime-size and limiter given e.g 'xs.15'
prime_size := cfg[0]
if prime_size !in primes {
return p_list
}
mut prime_count := extract_count(cfg[1])
if prime_count == 0 {
return primes[prime_size]
}
mut num := ''
for prime_count != 0 {
num = primes[prime_size][rand.int_in_range(0, primes[prime_size].len - 1)]
if num in p_list {
continue
}
p_list << num
prime_count -= 1
if prime_count >= primes[prime_size].len {
p_list = primes[prime_size]
break
}
} // eo-for
return p_list
} // cfg not understood
else {
return p_list
}
}
return p_list
}
pub fn random_set(cfg PrimeCfg) ?[]PrimeSet {
p_lists := [
cfg.r.split('.'),
cfg.a.split('.'),
cfg.b.split('.'),
].map(random_list(it.map(it.trim_space().to_lower())))
// test for empty lists
//
if p_lists.any(it.len == 0) {
msg := [
'bad config was :\n\n"$cfg"',
"maybe try e.g { r: 'l.5' a: 'xxl.5 b: 'xl.10' } makes a set of 250",
'your config was { $cfg.short() }',
'sizes $usage()\n',
].join('\n\n')
return error(msg)
}
// filter unique combinations thru map
//
mut tmp := map[string]PrimeSet{}
for r in p_lists[0] {
for a in p_lists[1] {
for b in p_lists[2] {
d := PrimeSet{r, a, b}
if d.key() in tmp {
continue
}
tmp[d.key()] = d
}
}
}
return tmp.keys().map(tmp[it])
}

View File

@ -0,0 +1,48 @@
[primes]
xs = [
"3", "5", "7", "11", "13", "17", "19", "21", "23", "27", "29", "31", "37", "41", "43"
]
s = [
"7823", "5419", "1889", "7213", "9817", "6521", "3299", "3709", "4217", "5737", "8753", "9187", "7013", "7951", "8647", "2089", "9241", "4993", "2819", "6469", "7879", "2341", "4733", "3259", "9697", "4373", "4231", "3001", "6577", "6449", "8783", "3947", "7687", "3221", "3449", "5527", "7411", "4339", "9461", "2909", "5449", "5843", "9007", "9461", "5419", "3571", "4133", "5417", "9043"
]
m = [
"55399", "57287", "53609", "13397", "53051", "95009", "44257", "15263", "47149", "53791", "67219", "35339", "92789", "40751", "27827", "89867", "45953", "81041", "34847", "66359", "73291", "31397", "19577", "31081", "73613", "57413", "42331", "61057", "68399", "21937", "15559", "31667", "39161", "12011", "32237", "28411", "43207", "17957", "60611", "61781", "61363", "92779", "87671", "68711", "59419"
]
ml = [
"3197239", "4777723", "5634029", "5345227", "3775663", "2577733", "2484731", "9489157", "9795311", "8807549", "4211929", "2192143", "1790611", "6124379", "7110449", "3423877", "3104219", "7823597", "4007569", "9538877", "9973409", "1337753", "2981431", "2052977", "3881461", "3107569", "6681733", "7208623", "6121067", "7024543", "7781927", "1355129", "3668941", "6326527", "8262383", "9656587", "8881111", "8026561", "8561011", "2398367", "8479967", "9903059"
]
l = [
"912325499", "724443641", "385137293", "926358929", "890097763", "130750339", "116017079", "144416561", "407856209", "410030717", "197927203", "202551331", "977001847", "538242283", "257127457", "818698583", "752243881", "596336339", "557356187", "240508001", "387269011", "456025643", "877258717", "369868633", "362001583", "864195281", "599680687", "215741453", "804623063", "938568131", "404699117", "255340313", "888734827"
]
xl = [
"4574666035183", "8417547875059", "4570443646591", "5048274773407", "7164005295869", "4518845199877", "5614561710673", "5260860119369", "7543626839449", "7276541368879", "7214431178011"
]
xxl = [
"835105110050347323134847370069", "143404874801173091148814190773", "201127251710413955312944973641", "404730213474815811826728006169", "667169667959358791072930585279", "501481765763667164242641192419", "446412154680934394834729305967", "460686874458826922684722184083", "830222457872535605976269306419", "185066433124965184224530740781"
]
xxxl = [
"41538632149409670203875645951895100583020903271493", "55019010857476970540160613582574368926321997435229", "80761049172599658440341257254896253118458575469719", "92617063636548016967671406179456164264446680329889", "40758128168942353395091585296464433019876977918493", "17036082713942833563217660171659128981053237307867"
]
mega = [
"198661663026514394979566031544326159107541699547309448720761733220478761329",
"992115042569270219928946682049053165598301867001123059433315541401938225427",
"101119309563131407758796252580471889739272930201455389991048893945940402311",
"516443083987514486594464741392901709033855328978282817026554473797439114541",
"660494613626358425463317297108260219544493771347620531225987686364487648369",
"175165388963515374726669791577365607804663690256056516815884117391457090227",
"121262978328120003250427616443327736340111055737282325346316041430003976093",
"532555459206091912293747167640580582656170020402363753480607569083534637187",
"839484247168370590833838270348781593651940895434421321502494070475339668381"
]
giga = [
"2884936882895429808357269729604622214585402525045525076838045027770319174012257013842658348401917179",
"8037083210794932674709412895115006556909092833822611837764114152389681777593745417056345965060910227",
"1910357352360710783480406392361973438148872596091183337139932590856337321762535807846055642301808061"
]
crazy = [
"351658062438469364437012934379378697325167334523785230519867682144348475237641332775452277714957257653189294629411183788781622118310946477143255873639",
"418013511717417121492936230539394143156615405871395138061025561400693302897049176138915109523457335116963751456368324397935946348902043658050622087259"
]
biggest = [
"477465885805093403826364314122183623487417165049378763285130618473465816825613056800831031290901922120098722325692953433392398177611894890290680254055832803915973853297598146797960159348700184977452096596048789018857150992316047149695774005795966830209155566710937235622005570209535860865999594857311"
]