SYCL: Add support for FLOOR,CEIL,ROUND and TRUNC unary operators

[safranowith]

This PR adds support for the following unary operators in the SYCL backend:

FLOOR, CEIL, ROUND, TRUNC.

Changes include:

Implementation in element_wise.cpp and element_wise.hpp

Header updates in ggml.h

Registration in SYCL.csv

Documentation updates in docs/ops.md

Test coverage in test-backend-ops.cpp

This SYCL implementation was added following the prior addition of the same operators in the CPU backend.

[CISC]

There's a bad merge or something, you are touching a lot of unrelated code.

[safranowith]

Thank you for the feedback! You're absolutely right — it looks like some unrelated changes ended up in this PR, likely due to an incorrect merge or leftover WIP. I'll clean it up and rebase the branch to ensure it only includes the intended SYCL support for floor, ceil, round, and trunc. ‫בתאריך יום ה׳, 16 באוק׳ 2025 ב-14:23 מאת ‪Sigbjørn Skjæret‬‏ <‪ ***@***.***‬‏>:‬

…

*CISC* left a comment (ggml-org/llama.cpp#16613) <#16613 (comment)> There's a bad merge or something, you are touching a lot of unrelated code. — Reply to this email directly, view it on GitHub <#16613 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BT72BRODH2N23TYYM4LQIK33X552BAVCNFSM6AAAAACJLNIRJ6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTIMJQGM4TSNRXGE> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[safranowith]

I've cleaned up the unrelated changes — let me know if everything looks good now or if there's anything else I should fix. ‫בתאריך יום ה׳, 16 באוק׳ 2025 ב-15:09 מאת ‪Sigbjørn Skjæret‬‏ <‪ ***@***.***‬‏>:‬

…

***@***.**** commented on this pull request. ------------------------------ In ggml/include/ggml.h <#16613 (comment)>: > @@ -1462,7 +1494,6 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b); - and here ------------------------------ In ggml/include/ggml.h <#16613 (comment)>: > @@ -1941,7 +1972,6 @@ extern "C" { int p1, // padding dimension 1 int d0, // dilation dimension 0 int d1); // dilation dimension 1 - and here ------------------------------ In ggml/include/ggml.h <#16613 (comment)>: > @@ -1976,7 +2006,6 @@ extern "C" { int d1, // dilation height int d2 // dilation depth ); - and here ------------------------------ In ggml/include/ggml.h <#16613 (comment)>: > - + and here — Reply to this email directly, view it on GitHub <#16613 (review)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BT72BRPLMUUIAIBI425B4ED3X6DHFAVCNFSM6AAAAACJLNIRJ6VHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMZTGNBUGQ4TCNRVG4> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[CISC]

Just a few more whitespace cleanups.

[CISC]

Let's try again. :)

[safranowith]

Is everything in order now? Please let me know if any further changes are needed. ‫בתאריך יום ה׳, 16 באוק׳ 2025 ב-15:41 מאת ‪Sigbjørn Skjæret‬‏ <‪ ***@***.***‬‏>:‬

…

***@***.**** commented on this pull request. Let's try again. :) ------------------------------ In ggml/include/ggml.h <#16613 (comment)>: > + GGML_API struct ggml_tensor * ggml_pad_ext( struct ggml_context * ctx, struct ggml_tensor * a, + int lp0, ⬇️ Suggested change - - GGML_API struct ggml_tensor * ggml_pad_ext( - struct ggml_context * ctx, - struct ggml_tensor * a, - - int lp0, + + GGML_API struct ggml_tensor * ggml_pad_ext( + struct ggml_context * ctx, + struct ggml_tensor * a, + int lp0, — Reply to this email directly, view it on GitHub <#16613 (review)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BT72BRKOZBOQ5JEFNJLGXYT3X6HAPAVCNFSM6AAAAACJLNIRJ6VHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMZTGNBUGYZTSMZVG4> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[NeoZhangJianyu]

It's great work!
It's first time that SYCL backend support OP ahead of other backend.

Thank you!

[CISC]

Also, did you rebase, or is the ops.md CI comparing against master?

[safranowith]

I've rebased on top of latest master and regenerated ops.md using create_ops_docs.py. Waiting for CI to complete — please let me know if anything else is needed! ‫בתאריך יום א׳, 19 באוק׳ 2025 ב-12:24 מאת ‪Sigbjørn Skjæret‬‏ <‪ ***@***.***‬‏>:‬

…

***@***.**** commented on this pull request. Also, did you rebase, or is the ops.md CI comparing against master? ------------------------------ In ggml/include/ggml.h <#16613 (comment)>: > + // GELU using erf (error function) when possible ⬇️ Suggested change - - // GELU using erf (error function) when possible + // GELU using erf (error function) when possible — Reply to this email directly, view it on GitHub <#16613 (review)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BT72BRM42VN7I6HO2A4YCCL3YNKFNAVCNFSM6AAAAACJLNIRJ6VHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMZTGNJUGM3TANRZGU> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[CISC]

I've rebased on top of latest master and regenerated ops.md using create_ops_docs.py.

You got it the other way around, you need to update your CSVs, Vulkan supports SSM_SCAN/CONV now.

[safranowith]

Thanks for the clarification — CSV and docs are now synced.
Is the PR ready for merge, or is anything else still needed?

[CISC]

Looks good to merge once @ggerganov approves.

Fixed the reason ops.md got out of sync in #16663

Edit: The reason is that #16463 got merged without updated CSV.