use tg declarations to handle mat and vec dimensions

[ZuseZ4]

No description provided.

[ZuseZ4]

@wsmoses are we sure that this is correct?

Enzyme/enzyme/Enzyme/Utils.cpp

Line 2103 in c5a6954

bool is_normal(llvm::IRBuilder<> &B, llvm::Value *trans) {

I just added handling for transposed matrices and my tests indicate that it's not.
I did modify the gemm_f_c_transpose_lacpy.ll testcase by setting transa to 'T' instead of 'N' (and A is modified, thus cached). In the T case the lacpy args to cache A should be switched as compared to the old 'N" testcase, but that's not happening.

Edit. sry, nvm. mixed up tests. Caching now works correct under transpositions.

[wsmoses]

That code is wrong there should never be a builder nor any load or store generated

…

On Thu, Jun 15, 2023 at 8:00 AM Manuel Drehwald ***@***.***> wrote: @wsmoses <https://github.com/wsmoses> are we sure that this is correct? https://github.com/EnzymeAD/Enzyme/blob/c5a6954570b6929ac10cdd7f6c8b3c505ecab82c/enzyme/Enzyme/Utils.cpp#L2103 I just added handling for transposed matrices and my tests indicate that it's not. I did modify the gemm_f_c_transpose_lacpy.ll testcase by setting transa to 'T' instead of 'N' (and A is modified, thus cached). — Reply to this email directly, view it on GitHub <#1281 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJTUXFVFNNEUMJSLAHLWZLXLKQGLANCNFSM6AAAAAAZHI5GZM> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[wsmoses]

The condition at the end should be return false at the end of the function not return true

…

On Thu, Jun 15, 2023 at 8:04 AM Billy Moses ***@***.***> wrote: That code is wrong there should never be a builder nor any load or store generated On Thu, Jun 15, 2023 at 8:00 AM Manuel Drehwald ***@***.***> wrote: > @wsmoses <https://github.com/wsmoses> are we sure that this is correct? > > https://github.com/EnzymeAD/Enzyme/blob/c5a6954570b6929ac10cdd7f6c8b3c505ecab82c/enzyme/Enzyme/Utils.cpp#L2103 > I just added handling for transposed matrices and my tests indicate that > it's not. > I did modify the gemm_f_c_transpose_lacpy.ll testcase by setting transa > to 'T' instead of 'N' (and A is modified, thus cached). > > — > Reply to this email directly, view it on GitHub > <#1281 (comment)>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/AAJTUXFVFNNEUMJSLAHLWZLXLKQGLANCNFSM6AAAAAAZHI5GZM> > . > You are receiving this because you were mentioned.Message ID: > ***@***.***> >

[wsmoses]

There should also be a continue after the for loop adding all the users into todo

…

On Thu, Jun 15, 2023 at 8:06 AM Billy Moses ***@***.***> wrote: The condition at the end should be return false at the end of the function not return true On Thu, Jun 15, 2023 at 8:04 AM Billy Moses ***@***.***> wrote: > That code is wrong there should never be a builder nor any load or store > generated > > On Thu, Jun 15, 2023 at 8:00 AM Manuel Drehwald ***@***.***> > wrote: > >> @wsmoses <https://github.com/wsmoses> are we sure that this is correct? >> >> https://github.com/EnzymeAD/Enzyme/blob/c5a6954570b6929ac10cdd7f6c8b3c505ecab82c/enzyme/Enzyme/Utils.cpp#L2103 >> I just added handling for transposed matrices and my tests indicate that >> it's not. >> I did modify the gemm_f_c_transpose_lacpy.ll testcase by setting transa >> to 'T' instead of 'N' (and A is modified, thus cached). >> >> — >> Reply to this email directly, view it on GitHub >> <#1281 (comment)>, >> or unsubscribe >> <https://github.com/notifications/unsubscribe-auth/AAJTUXFVFNNEUMJSLAHLWZLXLKQGLANCNFSM6AAAAAAZHI5GZM> >> . >> You are receiving this because you were mentioned.Message ID: >> ***@***.***> >> >

[ZuseZ4]

@wsmoses thanks. Behaves better now.

[ZuseZ4]

This now also has enough to remove this one: (ld $A, $transa, $lda, $m, $n),,
but since we already solved that I'll focus on nested rules before cleaning working code up.

[tgymnich]

this broke both macOS build and macOS tests

[ZuseZ4]

weren't they broken already for a while @tgymnich ?
Ah I see, they got fixed in between.

[tgymnich]

Yeah i fixed them last WE. The Xcode build is the only one still broken (but thats just the cmake arg order).