7 piece syzygy TB support #557
Conversation
support 7 man syzygy and 16 bit dtz flag
|
for some reason this seems to be a regression? I think it has to be another patch in next... |
|
Have you tested it with this patch but with the same tablebases on both sides? (To rule out tablebase performance issues?) |
|
Well that was 6 vs 7 and there were only 32 7 piece hits... So I think
that's close enough to the same.
…On Tue, May 8, 2018, 6:22 AM Tilps ***@***.***> wrote:
Have you tested it with this patch but with the same tablebases on both
sides? (To rule out tablebase performance issues?)
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
<#557 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AO6INLOsA-EGaTRXJawA0PG-94A2-TFtks5twXHmgaJpZM4T111Z>
.
|
|
last test had tilps nn cache difference. I think this would be safe to commit without regression... it may be a small regression, but I think it is unlikely, since it started out positive at first. Any regression is probably due to 7 man being on HD instead of SSD like my 6 man, and the fact that I only have not useful ones really. |
| @@ -501,7 +501,7 @@ int decompress_pairs(PairsData* d, uint64_t idx) { | |||
| uint32_t k = idx / d->span; | |||
|
|
|||
| // Then we read the corresponding SparseIndex[] entry | |||
| uint32_t block = number<uint32_t, LittleEndian>(&d->sparseIndex[k].block); | |||
| size_t block = number<uint32_t, LittleEndian>(&d->sparseIndex[k].block); | |||
| @@ -155,7 +155,7 @@ struct LR { | |||
| Sym get() { | |||
| return S == Left ? ((lr[1] & 0xF) << 8) | lr[0] : | |||
| S == Right ? (lr[2] << 4) | (lr[1] >> 4) : | |||
| S == Value ? lr[0] : (assert(false), Sym(-1)); | |||
| S == Value ? ((lr[1] & 0xF) << 8) | lr[0] : (assert(false), Sym(-1)); | |||
There was a problem hiding this comment.
Update the comment above to indicate that value can also be 12 bits now?
| if (flags & TBFlag::Mapped) | ||
| value = map[idx[WDLMap[wdl + 2]] + value]; | ||
| if (flags & TBFlag::Mapped) { | ||
| if (flags & TBFlag::LongDTZ) { |
There was a problem hiding this comment.
Indentation is a bit messed up here and the following few lines.
| for (int i = 0; i < 4; ++i) { // Sequence like 3,x,x,x,1,x,0,2,x,x | ||
| e.get(0, f)->map_idx[i] = (uint16_t)(data - e.map + 1); | ||
| data += *data + 1; | ||
| int flags = e.get(0, f)->flags; |
There was a problem hiding this comment.
indentation wrong on this line.
|
While I think it is cool, I wonder how many people will have 20TB of tablebases... I should also add that many many GUIs take care of this themselves, with no need for code from the engine. I presume Leela is not using the TBs within its search. |
|
No GUI does 7 piece syzygy currently, and some common ones don't support syzygy at all. Actually, cutechess is the only one I know that does. And there is no reason to comment on a PR if you don't bother to understand it enough to know what it does. Leela does use TB within search... And if you are only interested in analyzing one position, you only need that one 7 piece syzygy file, or a few that an 8 piece position might convert to, so you could use it with only a few GB of files. Even all the pawnless files all fit on my 4TB hard drive with room to store other things too. But there is no reason to limit an engine because few people would use it, see TCEC. |
|
Huh? ChessBase and Fritz have supported Syzygy for *years*, so, I think you are a bit uninformed.
…On Fri, May 18, 2018 at 12:14 PM, jjoshua2 ***@***.***> wrote:
No GUI does 7 piece syzygy currently, and common ones like
Chessbase/Fritz/Arena don't support syzygy at all. Actually, cutechess is
the only one I know that does.
And there is no reason to comment on a PR if you don't bother to
understand it enough to know what it does. Leela does use TB within
search...
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#557 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ADbG10JlYMW7e959gyIH-S7DXy7dMmMNks5tzuVWgaJpZM4T111Z>
.
|
|
I haven't used a Chessbase GUI for a few years, but I've used Arena and InfinityChess and Cutechess recently. Mostly cutechess. But this whole GUI thing is irrelevant, because the engine needs them for search and for root filtering. |
|
Then you could simply have said it uses the syzygy within its search, instead of telling me (pretty rudely), editor of ChessBase News for 8 years, CB and Fritz cannot use them. Especially in view of the fact that CB sells more than one syzygy package in their store. |
|
I apologize for the error about CB/Fritz. I didn't know you were associated with CB either, so easier to take an offense on your product. I mean no disrepect to it, although we are not going to optimize Leela for one particular GUI. I did clearly state that it is used in search, and if you had looked at the code you would have seen that before I even told you. This area is for code comments. If you are interested in more general things, please use the issues section. |
|
This whole PR might be irrelavent anyway since it's been a while without being merged, and focus is supposed to be on lc0 cudann now, which won't easily support syzygy. But for those that want it, they can use this build in the meantime since it is published now. |
support 7 man syzygy and 16 bit dtz flag