Add preprocessing TransposeExtractConcat pass #17692
Conversation
IanWood1
force-pushed
the
concat
branch
2 times, most recently
from
f256330 to
50d192f
Compare
compiler/src/iree/compiler/Preprocessing/Common/TransposeExtractConcat.cpp
Show resolved
Hide resolved
| rewriter, genericOp.getDpsInitOperand(0)->get(), perm); | ||
| SmallVector<Value> newOperands(genericOp.getOperands()); | ||
| newOperands.back() = newInit; | ||
| auto newGenericOp = |
There was a problem hiding this comment.
I would have expected that you have to transpose the operands as well? I am actually confused why you even need to do a transpose or anything here... you should be able to just clone the operation as is?
There was a problem hiding this comment.
I guess the naming for the function createTransposedInit wasn't good. Its not transposing anything but rather creating an empty tensor of the same shape as the new. The idea is that there is a long chain of unary elementwise generic ops and when the tensor.extract_slice op gets transposed, the corresponding child linalg generic's operand type is updated but not the dps init operand.
e.g.
generic with ins(tensor<1024x2>) outs(tensor<1024x2>)
then, after changing transposing extract slice ins and outs don't match:
generic with ins(tensor<2x1024>) outs(tensor<1024x2>)
There was a problem hiding this comment.
Lets try to find time today to walk through this a bit.
compiler/src/iree/compiler/Preprocessing/Common/TransposeExtractConcat.cpp
Outdated
Show resolved
Hide resolved
Signed-off-by: Ian Wood <ianwood2024@u.northwestern.edu>
|
Talked with Mahesh yesterday about this. Notes:
|
nod-ai/SHARK-ModelDev#683