Integrate llvm 1_20_2025 #19740
Integrate llvm 1_20_2025 #19740
Conversation
Signed-off-by: Nirvedh Meshram <[email protected]>
nirvedhmeshram
requested review from
hanhanW,
MaheshRavishankar,
pashu123 and
benvanik
as code owners
Signed-off-by: Nirvedh Meshram <[email protected]>
nirvedhmeshram
requested review from
antiagainst,
kuhar,
qedawkins and
Groverkss
as code owners
Signed-off-by: Nirvedh Meshram <[email protected]>
| // CHECK: arith.maxnumf | ||
| // CHECK: arith.maxnumf | ||
| // CHECK: vector.broadcast %{{.*}} : f32 to vector<4xf32> |
There was a problem hiding this comment.
I am not sure, that maxnumf is not present in IR, overall a lot of reduction codegen seems changed. I wonder if its something with valuebound interface like this commit
llvm/llvm-project#122804
I will flag this on discord.
There was a problem hiding this comment.
Actually I think whats happening here can be explained by this commit
llvm/llvm-project#118952
I believe that we were using -INF in softmax decompostion when we wanted NAN, with NAN, I think there is some simplification as you would simply pick the other argument hence needing one less maxnumf after folding..
There was a problem hiding this comment.
lets check what happens to correctness with this... Might need to revert it. I missed it, but I dont think using NAN make sense. Should be -INF.
There was a problem hiding this comment.
The CI is not finding any errors, I believe there are softmax dispactches checked into the regression tests so I think this works. I think there is some discussion on this here
llvm/llvm-project#114595
NaN is safe with maxnumf but should we have used maximumf and kept using -INF would be something to decide. Also should maxnumf fold the same for -INF is also something to think over.
Signed-off-by: Nirvedh Meshram <[email protected]>
nirvedhmeshram
force-pushed
the
integrate_1_20_2025
branch
from
a9ad59b to
d07a2ec
Compare
compiler/src/iree/compiler/Codegen/Common/DecomposePackUnPackOps.cpp
Outdated
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Common/TensorToVectorVectorizePad.cpp
Outdated
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Common/TensorToVectorVectorizePad.cpp
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/SPIRV/SPIRVVectorizeLoadStore.cpp
Outdated
Show resolved
Hide resolved
Signed-off-by: Nirvedh Meshram <[email protected]>
compiler/src/iree/compiler/Codegen/Common/DecomposePackUnPackOps.cpp
Outdated
Show resolved
Hide resolved
compiler/src/iree/compiler/Codegen/Common/DecomposePackUnPackOps.cpp
Outdated
Show resolved
Hide resolved
Signed-off-by: Nirvedh Meshram <[email protected]>
|
@MaheshRavishankar @kuhar let me know if you think I should land this? the only not discussed topic are the SPIRV test changes, seems like some value bounds improvements to me in them, but I don't read those shuffle ops too well. |
|
If tests pass, land it |
|
drive by nits for PR/branch naming: Pushing to the shared repository with a branch name matching If you use megabump (https://github.com/iree-org/megabump), the branch name will be generated for you: https://github.com/iree-org/megabump/blob/37cbfaf83374f51d6d02528745711d9d7d515a69/scripts/start_integrate#L16 |
Reverts
RISC-V backend
Python related changes
NVPTX changes
MLIR API changes (this change is breaking something in HLO)
TOSA (tosa.tile operation is changed and torch-mlir needs to catch up)
Updates to Torch-MLIR
floating type changes to be upstreamed
https://github.com/iree-org/torch-mlir/tree/fix_forward_iree_llvm_integrate_2025
style change