comic-hasher/hashing.go
Timmy Welch 0928ed6ccf Optimize memory usage
Add a basic map storage that does manual searches to conserve memory
Change saved hashes format to allow multiple hashes for a given ID
Add a vptree storage

Maps in Go take up a huge amount of space changing IDList to []ID took
  memory from over 1GB down to 200MB (note this was on aarch64 MacOS
  which for some reason uses less memory than aarch64 Linux).
  Exhaustive searches using slices took about 30 ms search now takes
  50-60 ms as it takes longer to iterate a map. Partial hashes will
  speed up searches to 8 ms at the cost of 700MB initial memory usage
  and 400MB idle (though this is on MacOS, which for some reason uses
  less memory that aarch64 Linux so probably more like
  900MB initial -> 600 MB idle on an RPI running Linux)
2024-09-07 14:51:18 -07:00

249 lines
5.0 KiB
Go

package ch
import (
"cmp"
_ "embed"
"fmt"
"image"
"log"
"math/bits"
"runtime"
"slices"
"gitea.narnian.us/lordwelch/goimagehash"
)
//go:embed hashes.gz
var Hashes []byte
const (
H0 uint64 = 0b11111111 << (8 * iota)
H1
H2
H3
H4
H5
H6
H7
)
const (
Shift0 = (8 * iota)
Shift1
Shift2
Shift3
Shift4
Shift5
Shift6
Shift7
)
const (
ComicVine Source = "comicvine.gamespot.com"
)
type Source string
type Match struct {
Distance int
Hash uint64
}
type ID struct {
Domain Source
ID string
}
type Result struct {
IDs IDList
Distance int
Hash Hash
}
type Im struct {
Im image.Image
Format string
Path string
ID ID
}
type ImageHash struct {
Hashes []Hash
ID ID
}
type Hash struct {
Hash uint64
Kind goimagehash.Kind
}
// IDList is a map of domain to ID eg IDs["comicvine.gamespot.com"] = []string{"1235"}
// Maps are extremely expensive in go for small maps this should only be used to return info to a user no internal code should use this
type IDList map[Source][]string
type OldSavedHashes map[Source]map[string][3]uint64
type SavedHashes struct {
IDs [][]ID
Hashes [3]map[uint64]int
}
func ToIDList(ids []ID) IDList {
idlist := IDList{}
for _, id := range ids {
idlist[id.Domain] = Insert(idlist[id.Domain], id.ID)
}
return idlist
}
func InsertID(ids []ID, id ID) []ID {
index, itemFound := slices.BinarySearchFunc(ids, id, func(e ID, t ID) int {
return cmp.Or(
cmp.Compare(e.Domain, t.Domain),
cmp.Compare(e.ID, t.ID),
)
})
if itemFound {
return ids
}
return slices.Insert(ids, index, id)
}
func (s *SavedHashes) InsertHash(hash Hash, id ID) {
for i, h := range s.Hashes {
if h == nil {
s.Hashes[i] = make(map[uint64]int)
}
}
hashType := int(hash.Kind) - 1
idx, hashFound := s.Hashes[hashType][hash.Hash]
if !hashFound {
idx = len(s.IDs)
s.IDs = append(s.IDs, make([]ID, 0, 3))
}
s.IDs[idx] = InsertID(s.IDs[idx], id)
s.Hashes[hashType][hash.Hash] = idx
}
func ConvertSavedHashes(oldHashes OldSavedHashes) SavedHashes {
t := SavedHashes{}
idcount := 0
for _, ids := range oldHashes {
idcount += len(ids)
}
t.IDs = make([][]ID, 0, idcount)
t.Hashes[0] = make(map[uint64]int, idcount)
t.Hashes[1] = make(map[uint64]int, idcount)
t.Hashes[2] = make(map[uint64]int, idcount)
for domain, sourceHashes := range oldHashes {
for id, hashes := range sourceHashes {
idx := len(t.IDs)
t.IDs = append(t.IDs, []ID{{domain, id}})
for hashType, hash := range hashes {
t.Hashes[hashType][hash] = idx
}
}
}
fmt.Println("Expected number of IDs", idcount)
idcount = 0
for _, ids := range t.IDs {
idcount += len(ids)
}
fmt.Println("length of hashes", len(t.Hashes[0])+len(t.Hashes[1])+len(t.Hashes[2]))
fmt.Println("Length of ID lists", len(t.IDs))
fmt.Println("Total number of IDs", idcount)
return t
}
type NewIDs struct {
OldID ID
NewID ID
}
type HashStorage interface {
GetMatches(hashes []Hash, max int, exactOnly bool) ([]Result, error)
MapHashes(ImageHash)
DecodeHashes(hashes SavedHashes) error
EncodeHashes() (SavedHashes, error)
AssociateIDs(newIDs []NewIDs)
GetIDs(id ID) IDList
}
func Atleast(maxDistance int, searchHash uint64, hashes []uint64) []Match {
matchingHashes := make([]Match, 0, len(hashes)/2) // hope that we don't need all of them
for _, storedHash := range hashes {
distance := bits.OnesCount64(searchHash ^ storedHash)
if distance <= maxDistance {
matchingHashes = append(matchingHashes, Match{distance, storedHash})
}
}
return matchingHashes
}
func Insert[S ~[]E, E cmp.Ordered](slice S, item E) S {
index, itemFound := slices.BinarySearch(slice, item)
if itemFound {
return slice
}
return slices.Insert(slice, index, item)
}
func InsertIdx[S ~[]E, E cmp.Ordered](slice S, item E) (S, int) {
index, itemFound := slices.BinarySearch(slice, item)
if itemFound {
return slice, index
}
return slices.Insert(slice, index, item), index
}
func MemStats() uint64 {
var m runtime.MemStats
runtime.ReadMemStats(&m)
return m.Alloc
}
func HashImage(i Im) ImageHash {
if i.Format == "webp" {
i.Im = goimagehash.FancyUpscale(i.Im.(*image.YCbCr))
}
var (
err error
)
ahash, err := goimagehash.AverageHash(i.Im)
if err != nil {
msg := fmt.Sprintf("Failed to ahash Image: %s", err)
log.Println(msg)
return ImageHash{}
}
dhash, err := goimagehash.DifferenceHash(i.Im)
if err != nil {
msg := fmt.Sprintf("Failed to dhash Image: %s", err)
log.Println(msg)
return ImageHash{}
}
phash, err := goimagehash.PerceptionHash(i.Im)
if err != nil {
msg := fmt.Sprintf("Failed to phash Image: %s", err)
log.Println(msg)
return ImageHash{}
}
return ImageHash{
Hashes: []Hash{{ahash.GetHash(), ahash.GetKind()}, {dhash.GetHash(), dhash.GetKind()}, {phash.GetHash(), phash.GetKind()}},
ID: i.ID,
}
}
func SplitHash(hash uint64) [8]uint8 {
return [8]uint8{
uint8((hash & H7) >> Shift7),
uint8((hash & H6) >> Shift6),
uint8((hash & H5) >> Shift5),
uint8((hash & H4) >> Shift4),
uint8((hash & H3) >> Shift3),
uint8((hash & H2) >> Shift2),
uint8((hash & H1) >> Shift1),
uint8((hash & H0) >> Shift0),
}
}