Why are the cold start runs faster, shouldn't it be the other way around?

It never really speeds up, it's always like 30% faster after I toggle warp timer on/off.

ArnoHu wrote:

Why are the cold start runs faster, shouldn't it be the other way around?

It never really speeds up, it's always like 30% faster after I toggle warp timer on/off.

OK, that's true, still it's the recompile that triggers the change. Optimizer can be wrong, or can disable optimizations for certain functions for one reason or another.

waabooboo wrote:

https://scratch.mit.edu/projects/1202302970/

Wolverine 3's official release. Some more small improvements have been made over the DCI version. Same interface as Wolverine 2, although I am in the process of designing a better UI

sprunkrLover1234 wrote:

How is your chess algorithm coming along? Just asking.

ArnoHu wrote:

Scratch chess engines TurbowWarp test game series, broadcasted live on lichess:

https://lichess.org/broadcast/scratch-chess-engine-test-games-series/round-1/SLvxmTkh#players

Featuring Wolverine 3, Black Crow, GoK Classic, GoK NNUE (1sec), Delta 5 and Stockfish 12 (depth 10).

GoK NNUE - Stockfish 12: 61 - 39
GoK NNUE - Black Crow: 72.5 - 27.5

GoK Classic - Wolverine 3: 63 - 37
GoK Classic - White Dove: 75.5 - 24.5

Currently playing:

GoK NNUE - Wolverine 3 NNUE
Black Crow - Stockfish 12

waabooboo wrote:

ArnoHu wrote:

Scratch chess engines TurbowWarp test game series, broadcasted live on lichess:

https://lichess.org/broadcast/scratch-chess-engine-test-games-series/round-1/SLvxmTkh#players

Featuring Wolverine 3, Black Crow, GoK Classic, GoK NNUE (1sec), Delta 5 and Stockfish 12 (depth 10).

GoK NNUE - Stockfish 12: 61 - 39
GoK NNUE - Black Crow: 72.5 - 27.5

GoK Classic - Wolverine 3: 63 - 37
GoK Classic - White Dove: 75.5 - 24.5

Currently playing:

GoK NNUE - Wolverine 3 NNUE
Black Crow - Stockfish 12

https://limewire.com/d/dSiUz#DNR14ViCEd

Try this – this one uses a Smallbrain net (not the 3 MB, one that's much smaller but still larger than the Leorik net).

I am curious how this does against GoK NNUE, as well as against the current Wolverine NNUE that uses the Leorik net. (I can't get two Wolverines running in a DCI project due to variable conflicts, so I can't make direct comparisons – but I suspect the Smallbrain version is stronger even though it's about 3x slower.)

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

Scratch chess engines TurbowWarp test game series, broadcasted live on lichess:

https://lichess.org/broadcast/scratch-chess-engine-test-games-series/round-1/SLvxmTkh#players

Featuring Wolverine 3, Black Crow, GoK Classic, GoK NNUE (1sec), Delta 5 and Stockfish 12 (depth 10).

GoK NNUE - Stockfish 12: 61 - 39
GoK NNUE - Black Crow: 72.5 - 27.5

GoK Classic - Wolverine 3: 63 - 37
GoK Classic - White Dove: 75.5 - 24.5

Currently playing:

GoK NNUE - Wolverine 3 NNUE
Black Crow - Stockfish 12

https://limewire.com/d/dSiUz#DNR14ViCEd

Try this – this one uses a Smallbrain net (not the 3 MB, one that's much smaller but still larger than the Leorik net).

I am curious how this does against GoK NNUE, as well as against the current Wolverine NNUE that uses the Leorik net. (I can't get two Wolverines running in a DCI project due to variable conflicts, so I can't make direct comparisons – but I suspect the Smallbrain version is stronger even though it's about 3x slower.)

Couldn't download it.

ArnoHu wrote:

waabooboo wrote:

https://scratch.mit.edu/projects/1202302970/

Wolverine 3's official release. Some more small improvements have been made over the DCI version. Same interface as Wolverine 2, although I am in the process of designing a better UI

Cool, congrats to this version, it is super dangerous!

Where are the DCI-enabled versions? Because I still have only those from a couple of weeks ago.

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

https://scratch.mit.edu/projects/1202302970/

Wolverine 3's official release. Some more small improvements have been made over the DCI version. Same interface as Wolverine 2, although I am in the process of designing a better UI

Cool, congrats to this version, it is super dangerous!

Where are the DCI-enabled versions? Because I still have only those from a couple of weeks ago.

The DCI file I shared is the new version, only change is I swapped out the Leorik net for a Smallbrain one. Again I had some issues sharing it as a Scratch project, that's why I shared it as a file. The second link I sent will hopefully work…

And I figured out what causes the major slowdowns after a long, painful process, testing every possible thing. Turns out it's sprite costumes with numbers as names, for whatever reason that slows things down significantly…

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

https://scratch.mit.edu/projects/1202302970/

Wolverine 3's official release. Some more small improvements have been made over the DCI version. Same interface as Wolverine 2, although I am in the process of designing a better UI

Cool, congrats to this version, it is super dangerous!

Where are the DCI-enabled versions? Because I still have only those from a couple of weeks ago.

The DCI file I shared is the new version, only change is I swapped out the Leorik net for a Smallbrain one. Again I had some issues sharing it as a Scratch project, that's why I shared it as a file. The second link I sent will hopefully work…

And I figured out what causes the major slowdowns after a long, painful process, testing every possible thing. Turns out it's sprite costumes with numbers as names, for whatever reason that slows things down significantly…

Great finding, I uncovered the same thing only a week ago, it's crazy!

Reason:
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L88-L99
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L150-L160

Why?
In Scratch, switch costume to (“1”) picks the costume named “1”, while switch costume to (1) picks costume number 1. Any stored value could later reach a switch-costume block. So when some costume or sound is named “1”, the compiler keeps every stored 1 as the string “1” to preserve that behaviour. Each later use in arithmetic or a comparison then needs a string-to-number conversion, and the compiler can't use its typed number code paths.

Without S3 backward-compatibility, engines would play a lot better.

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

https://scratch.mit.edu/projects/1202302970/

Wolverine 3's official release. Some more small improvements have been made over the DCI version. Same interface as Wolverine 2, although I am in the process of designing a better UI

Cool, congrats to this version, it is super dangerous!

Where are the DCI-enabled versions? Because I still have only those from a couple of weeks ago.

The DCI file I shared is the new version, only change is I swapped out the Leorik net for a Smallbrain one. Again I had some issues sharing it as a Scratch project, that's why I shared it as a file. The second link I sent will hopefully work…

And I figured out what causes the major slowdowns after a long, painful process, testing every possible thing. Turns out it's sprite costumes with numbers as names, for whatever reason that slows things down significantly…

Great finding, I uncovered the same thing only a week ago, it's crazy!

Reason:
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L88-L99
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L150-L160

Why?
In Scratch, switch costume to (“1”) picks the costume named “1”, while switch costume to (1) picks costume number 1. Any stored value could later reach a switch-costume block. So when some costume or sound is named “1”, the compiler keeps every stored 1 as the string “1” to preserve that behaviour. Each later use in arithmetic or a comparison then needs a string-to-number conversion, and the compiler can't use its typed number code paths.

Without S3 backward-compatibility, engines would play a lot better.

You should have told me, would have saved me some time… How does Wolverine with the Smallbrain net do, did you get a chance to test?

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

https://scratch.mit.edu/projects/1202302970/

Wolverine 3's official release. Some more small improvements have been made over the DCI version. Same interface as Wolverine 2, although I am in the process of designing a better UI

Cool, congrats to this version, it is super dangerous!

Where are the DCI-enabled versions? Because I still have only those from a couple of weeks ago.

The DCI file I shared is the new version, only change is I swapped out the Leorik net for a Smallbrain one. Again I had some issues sharing it as a Scratch project, that's why I shared it as a file. The second link I sent will hopefully work…

And I figured out what causes the major slowdowns after a long, painful process, testing every possible thing. Turns out it's sprite costumes with numbers as names, for whatever reason that slows things down significantly…

Great finding, I uncovered the same thing only a week ago, it's crazy!

Reason:
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L88-L99
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L150-L160

Why?
In Scratch, switch costume to (“1”) picks the costume named “1”, while switch costume to (1) picks costume number 1. Any stored value could later reach a switch-costume block. So when some costume or sound is named “1”, the compiler keeps every stored 1 as the string “1” to preserve that behaviour. Each later use in arithmetic or a comparison then needs a string-to-number conversion, and the compiler can't use its typed number code paths.

Without S3 backward-compatibility, engines would play a lot better.

You should have told me, would have saved me some time… How does Wolverine with the Smallbrain net do, did you get a chance to test?

Yeah I didn't know your costume names.

I played two quick rounds using Wolverine 3 NNUE-SB, 2.5 - 7.5 against GoK Classic, and 0 - 10 against GoK NNUE. I can export the games to lichess if you want.

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

https://scratch.mit.edu/projects/1202302970/

Wolverine 3's official release. Some more small improvements have been made over the DCI version. Same interface as Wolverine 2, although I am in the process of designing a better UI

Cool, congrats to this version, it is super dangerous!

Where are the DCI-enabled versions? Because I still have only those from a couple of weeks ago.

The DCI file I shared is the new version, only change is I swapped out the Leorik net for a Smallbrain one. Again I had some issues sharing it as a Scratch project, that's why I shared it as a file. The second link I sent will hopefully work…

And I figured out what causes the major slowdowns after a long, painful process, testing every possible thing. Turns out it's sprite costumes with numbers as names, for whatever reason that slows things down significantly…

Great finding, I uncovered the same thing only a week ago, it's crazy!

Reason:
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L88-L99
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L150-L160

Why?
In Scratch, switch costume to (“1”) picks the costume named “1”, while switch costume to (1) picks costume number 1. Any stored value could later reach a switch-costume block. So when some costume or sound is named “1”, the compiler keeps every stored 1 as the string “1” to preserve that behaviour. Each later use in arithmetic or a comparison then needs a string-to-number conversion, and the compiler can't use its typed number code paths.

Without S3 backward-compatibility, engines would play a lot better.

You should have told me, would have saved me some time… How does Wolverine with the Smallbrain net do, did you get a chance to test?

Yeah I didn't know your costume names.

I played two quick rounds using Wolverine 3 NNUE-SB, 2.5 - 7.5 against GoK Classic, and 0 - 10 against GoK NNUE. I can export the games to lichess if you want.

GoK Classic beat it 7.5-2.5, really? I wonder if I didn't set the NNUE up right, or maybe it was Wolverine HCE playing again…

0-10 against GoK NNUE is a bit disappointing, although not shocking. But I would like to see the exported games, especially against Classic.

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

https://scratch.mit.edu/projects/1202302970/

Wolverine 3's official release. Some more small improvements have been made over the DCI version. Same interface as Wolverine 2, although I am in the process of designing a better UI

Cool, congrats to this version, it is super dangerous!

Where are the DCI-enabled versions? Because I still have only those from a couple of weeks ago.

The DCI file I shared is the new version, only change is I swapped out the Leorik net for a Smallbrain one. Again I had some issues sharing it as a Scratch project, that's why I shared it as a file. The second link I sent will hopefully work…

And I figured out what causes the major slowdowns after a long, painful process, testing every possible thing. Turns out it's sprite costumes with numbers as names, for whatever reason that slows things down significantly…

Great finding, I uncovered the same thing only a week ago, it's crazy!

Reason:
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L88-L99
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L150-L160

Why?
In Scratch, switch costume to (“1”) picks the costume named “1”, while switch costume to (1) picks costume number 1. Any stored value could later reach a switch-costume block. So when some costume or sound is named “1”, the compiler keeps every stored 1 as the string “1” to preserve that behaviour. Each later use in arithmetic or a comparison then needs a string-to-number conversion, and the compiler can't use its typed number code paths.

Without S3 backward-compatibility, engines would play a lot better.

You should have told me, would have saved me some time… How does Wolverine with the Smallbrain net do, did you get a chance to test?

Yeah I didn't know your costume names.

I played two quick rounds using Wolverine 3 NNUE-SB, 2.5 - 7.5 against GoK Classic, and 0 - 10 against GoK NNUE. I can export the games to lichess if you want.

GoK Classic beat it 7.5-2.5, really? I wonder if I didn't set the NNUE up right, or maybe it was Wolverine HCE playing again…

0-10 against GoK NNUE is a bit disappointing, although not shocking. But I would like to see the exported games, especially against Classic.

What I can say is that the main Wolverine 3 NNUE clearly plays better. It won 6 - 4 against GoK Classic.

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

https://scratch.mit.edu/projects/1202302970/

Wolverine 3's official release. Some more small improvements have been made over the DCI version. Same interface as Wolverine 2, although I am in the process of designing a better UI

Cool, congrats to this version, it is super dangerous!

Where are the DCI-enabled versions? Because I still have only those from a couple of weeks ago.

The DCI file I shared is the new version, only change is I swapped out the Leorik net for a Smallbrain one. Again I had some issues sharing it as a Scratch project, that's why I shared it as a file. The second link I sent will hopefully work…

And I figured out what causes the major slowdowns after a long, painful process, testing every possible thing. Turns out it's sprite costumes with numbers as names, for whatever reason that slows things down significantly…

Great finding, I uncovered the same thing only a week ago, it's crazy!

Reason:
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L88-L99
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L150-L160

Why?
In Scratch, switch costume to (“1”) picks the costume named “1”, while switch costume to (1) picks costume number 1. Any stored value could later reach a switch-costume block. So when some costume or sound is named “1”, the compiler keeps every stored 1 as the string “1” to preserve that behaviour. Each later use in arithmetic or a comparison then needs a string-to-number conversion, and the compiler can't use its typed number code paths.

Without S3 backward-compatibility, engines would play a lot better.

You should have told me, would have saved me some time… How does Wolverine with the Smallbrain net do, did you get a chance to test?

Yeah I didn't know your costume names.

I played two quick rounds using Wolverine 3 NNUE-SB, 2.5 - 7.5 against GoK Classic, and 0 - 10 against GoK NNUE. I can export the games to lichess if you want.

GoK Classic beat it 7.5-2.5, really? I wonder if I didn't set the NNUE up right, or maybe it was Wolverine HCE playing again…

0-10 against GoK NNUE is a bit disappointing, although not shocking. But I would like to see the exported games, especially against Classic.

What I can say is that the main Wolverine 3 NNUE clearly plays better. It won 6 - 4 against GoK Classic.

Yeah that's fine, but I still want to try to diagnose the issue with the Smallbrain net – it really should be stronger than the Leorik one.

There is a larger Leorik net I can try as well, with 256-wide accumulators instead of 128. But I don't think we're seeing the full potential of the Smallbrain net, I'd like to figure that out first…

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

https://scratch.mit.edu/projects/1202302970/

Wolverine 3's official release. Some more small improvements have been made over the DCI version. Same interface as Wolverine 2, although I am in the process of designing a better UI

Cool, congrats to this version, it is super dangerous!

Where are the DCI-enabled versions? Because I still have only those from a couple of weeks ago.

The DCI file I shared is the new version, only change is I swapped out the Leorik net for a Smallbrain one. Again I had some issues sharing it as a Scratch project, that's why I shared it as a file. The second link I sent will hopefully work…

And I figured out what causes the major slowdowns after a long, painful process, testing every possible thing. Turns out it's sprite costumes with numbers as names, for whatever reason that slows things down significantly…

Great finding, I uncovered the same thing only a week ago, it's crazy!

Reason:
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L88-L99
https://github.com/TurboWarp/scratch-vm/blob/c4823421cb7c17d8d8a89878851ce1668c26a21f/src/compiler/irgen.js#L150-L160

Why?
In Scratch, switch costume to (“1”) picks the costume named “1”, while switch costume to (1) picks costume number 1. Any stored value could later reach a switch-costume block. So when some costume or sound is named “1”, the compiler keeps every stored 1 as the string “1” to preserve that behaviour. Each later use in arithmetic or a comparison then needs a string-to-number conversion, and the compiler can't use its typed number code paths.

Without S3 backward-compatibility, engines would play a lot better.

You should have told me, would have saved me some time… How does Wolverine with the Smallbrain net do, did you get a chance to test?

Yeah I didn't know your costume names.

I played two quick rounds using Wolverine 3 NNUE-SB, 2.5 - 7.5 against GoK Classic, and 0 - 10 against GoK NNUE. I can export the games to lichess if you want.

GoK Classic beat it 7.5-2.5, really? I wonder if I didn't set the NNUE up right, or maybe it was Wolverine HCE playing again…

0-10 against GoK NNUE is a bit disappointing, although not shocking. But I would like to see the exported games, especially against Classic.

What I can say is that the main Wolverine 3 NNUE clearly plays better. It won 6 - 4 against GoK Classic.

Yeah that's fine, but I still want to try to diagnose the issue with the Smallbrain net – it really should be stronger than the Leorik one.

There is a larger Leorik net I can try as well, with 256-wide accumulators instead of 128. But I don't think we're seeing the full potential of the Smallbrain net, I'd like to figure that out first…

I can run a JS profiler later and find out where the hotspot is.

Interestingly with 2sec think time GoK Classic does better against Wolverine 3 NNUE, 14 - 13 right now.