image.GetPoint() is extremely slow

[MTRNord]

Hi.

I am trying to reimplement the pdq hash algorithm in go and need to access all pixels and their r g b values individually for the whole image. It seems to work fine up until I call image.GetPoint. Doing this seems to significantly (speak SECONDS per iteration aka about 2 minutes for the whole file) slow down the process.

I am not sure why its slow.

The code used is like this:

       numCols := image.Width()
	numRows := image.Height()
	for i := 0; i < numRows; i++ {
		for j := 0; j < numCols; j++ {
			colorArray, err := image.GetPoint(j, i)
			if err != nil {
				log.Fatalf("Error getting pixel: %v", err)
			}
			r := colorArray[0]
			g := colorArray[1]
			b := colorArray[2]
			(*luma)[i*numCols+j] = LUMA_FROM_R_COEFF*r + LUMA_FROM_G_COEFF*g + LUMA_FROM_B_COEFF*b
		}
	}

where image is a reference to the image. The rest is semi relevant only. I verified that the GetPoint call specifically is the slow part of the loop.

Any ideas what might be going on or how to make it faster?

[MTRNord]

Converting via ToImage to the golang image.Image and then using At instead seems to make it work in milliseconds

[tonimelisma]

Hey @MTRNord, thanks for this. I don't have a clear answer unfortunately. Is it possible for you to try profiling the govips part?

[MTRNord]

Hi So I ran this now and I got this:

(I ran it on my original code)

➜  pdqhash-go git:(main) ✗ go tool pprof scanner q0004.prof 
File: scanner
Build ID: c82775b31f8a5d58fa8cb57bc7c7660cf278fc97
Type: cpu
Time: Feb 25, 2024 at 12:26pm (CET)
Duration: 57.12s, Total samples = 60.83s (106.49%)
Entering interactive mode (type "help" for commands, "o" for options)
(pprof) top10
Showing nodes accounting for 60s, 98.64% of 60.83s total
Dropped 89 nodes (cum <= 0.30s)
Showing top 10 nodes out of 26
      flat  flat%   sum%        cum   cum%
    36.54s 60.07% 60.07%     36.60s 60.17%  runtime.cgocall
    13.68s 22.49% 82.56%     13.68s 22.49%  [scanner]
     3.15s  5.18% 87.74%      3.15s  5.18%  [libc.so.6]
     2.91s  4.78% 92.52%      2.91s  4.78%  [libglib-2.0.so.0.7800.3]
     1.79s  2.94% 95.46%      1.79s  2.94%  [libgobject-2.0.so.0.7800.3]
     1.42s  2.33% 97.80%      1.42s  2.33%  [libvips.so.42.17.1]
     0.34s  0.56% 98.36%      0.34s  0.56%  runtime/internal/syscall.Syscall6
     0.14s  0.23% 98.59%     22.66s 37.25%  runtime._ExternalCode
     0.02s 0.033% 98.62%      0.39s  0.64%  internal/poll.(*FD).Write
     0.01s 0.016% 98.64%     37.50s 61.65%  github.com/MTRNord/pdqhash-go.(*PDQHasher).fillFloatLumaFromBufferImage

so the FFI seems to be either quite expensive or libvips is here.

Libvips was started like this:

vips.LoggingSettings(nil, vips.LogLevelMessage)
	vips.Startup(&vips.Config{
		ConcurrencyLevel: 0,
		MaxCacheFiles:    5,
		MaxCacheMem:      50 * 1024 * 1024,
		MaxCacheSize:     100,
		ReportLeaks:      false,
		CacheTrace:       false,
		CollectStats:     false,
	})
	defer vips.Shutdown()

The profile is at https://github.com/MTRNord/pdqhash-go/blob/2290c5881ff302dc009ac741b92a8682a39c1a38/q0004.prof

The profile was made based on the commit at https://github.com/MTRNord/pdqhash-go/tree/2290c5881ff302dc009ac741b92a8682a39c1a38 with scanner being build using CC=clang CCX=clang++ go build -asan cmd/scanner.go and ran using ./scanner -folder test-images/reg-test-input/labelme-subset/q0004.jpg -cpuprofile=q0004.prof

I hope that helps and if not I am happy to add more info :)

[MTRNord]

Looking at this I wonder if golang/go#19574 and similar issues are related here possibly.

[MTRNord]

Oh and the problematic call is at https://github.com/MTRNord/pdqhash-go/blob/2290c5881ff302dc009ac741b92a8682a39c1a38/pdq_hasher.go#L160 which is called for every pixel in the image.

[MTRNord]

Converting via ToImage to the golang image.Image and then using At instead seems to make it work in milliseconds

While this works in general I realised that this will introduce some slight (compression? quality?) error which in my case makes it fail as hashes are not generated as they should be :)