[aievec] Add lowering for vector.extract_strided_slice #491
Conversation
jsetoain
force-pushed
the
lower-extract-strided-slice
branch
2 times, most recently
from
e90d5ac
to
9e5589b
Compare
jsetoain
force-pushed
the
lower-extract-strided-slice
branch
from
9e5589b
to
e85f4ee
Compare
jsetoain
force-pushed
the
lower-extract-strided-slice
branch
3 times, most recently
from
422b500
to
6fe20bc
Compare
jsetoain
force-pushed
the
lower-extract-strided-slice
branch
2 times, most recently
from
b32514f
to
0179d9b
Compare
jsetoain
force-pushed
the
lower-extract-strided-slice
branch
4 times, most recently
from
ed13e90
to
371021f
Compare
jsetoain
force-pushed
the
lower-extract-strided-slice
branch
6 times, most recently
from
0a9563a
to
b9ec2c0
Compare
There was a problem hiding this comment.
Hi @jsetoain:
This looks very good. I would like @linay-xsj to go through the changes in lib/Dialect/AIEVec/Transforms/FoldMulAddChainToConvOp.cpp and
lib/Dialect/AIEVec/Transforms/FoldMulAddChainToConvOp.cpp
in detail. Thanks!
| printf("%d]\n", buff[nlanes - 1]); | ||
| } | ||
|
|
||
| void printv16xi32(v16int32 v) { printv<16, int32_t>(v); } |
There was a problem hiding this comment.
The aie_api.h already has a bunch of nice utilities to print vectors, for example, to print a v16int32, I could call:
aie::print(aie::vector<uint32_t,16>(myV16int32),true, "myV16int32: ");
There was a problem hiding this comment.
I had forgotten about these. I've switched to aie api to print vectors, and I've also added missing checks for unaligned loads in aieml.
jsetoain
force-pushed
the
lower-extract-strided-slice
branch
2 times, most recently
from
57f18a1
to
77fcf18
Compare
This is part of the canonical lowering of unaligned transfer read ops. For now we only support extracting a contiguous vector of half the size of the source vector.
jsetoain
force-pushed
the
lower-extract-strided-slice
branch
from
77fcf18
to
0a5d164
Compare
I think it's best if we do that after landing the patch. It's a bit messy, and it risks becoming even more of a liability. I've already got a couple more patching that will help with clean-up. |
This patch adds the lowering for
vector.extract_strided_sliceas a canonical solution to unaligned loads. This process breaks convolution detection and optimization for AIE-ML, and this patch addresses those as well.