[AIE2] Combiners for 8x8->8x8 and 8x4->4x8 matrix transposes #76
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
ValentijnvdBeek
added
llvm:globalisel
ValentijnvdBeek
requested review from
abhinay-anubola,
abnikant,
andcarminati,
gbossu,
khallouh,
konstantinschwarz,
martien-de-jong,
SagarMaheshwari99 and
stephenneuendorffer
as code owners
konstantinschwarz
requested changes
Jun 21, 2024
llvm/test/CodeGen/AIE/aie2/GlobalISel/prelegalizercombiner-shufflevector.mir
Outdated
Show resolved
Hide resolved
llvm/test/CodeGen/AIE/aie2/GlobalISel/prelegalizercombiner-shufflevector.mir
Outdated
Show resolved
Hide resolved
| %1:_(<64 x s8>) = COPY $x0 | ||
| %2:_(<64 x s8>) = COPY $x1 | ||
| %0:_(<64 x s8>) = G_SHUFFLE_VECTOR %1:_(<64 x s8>), %2:_, shufflemask(0, 16, 32, 48, 1, 17, 33, 49, 2, 18, 34, 50, 3, 19, 35, 51, 4, 20, 36, 52, 5, 21, 37, 53, 6, 22, 38, 54, 7, 23, 39, 55, 8, 24, 40, 56, 9, 25, 41, 57, 10, 26, 42, 58, 11, 27, 43, 59, 12, 28, 44, 60, 13, 29, 45, 61, 14, 30, 46, 62, 15, 31, 47, 63) |
There was a problem hiding this comment.
Mode 35 operates on a 8-bit element 8x8 matrix.
That would need to match shufflemask(0, 8, 16, 24, 1, 9, 17, 25, ...)?
There was a problem hiding this comment.
Yes, it does but vshuffle takes 1024 bytes of input. It ignores the higher bits of the input. 35 operates on a 8x8 8-bit element matrix, 64x8 vector, which is 512-bits. In the first case, those two are split into two 4x8 8-bit vectors and the second case is the "common" case where we just ignore the higher order bits.
What you propose would be a 4x8 8-bit match which is a 32x8 vector or 256-bit.
ValentijnvdBeek
force-pushed
the
vvandebe.shufflevector.pattern.optimization
branch
from
0e664aa to
f872cf4
Compare
Yeah, it is a strange set of the ISA. At the moment I am orientating myself by looking at the image descriptions since from the perspective of |
ValentijnvdBeek
force-pushed
the
vvandebe.shufflevector.pattern.optimization
branch
from
f872cf4 to
73a92c2
Compare
ValentijnvdBeek
changed the title
ValentijnvdBeek
force-pushed
the
vvandebe.vshuffle.impl
branch
from
be3751a to
5654047
Compare
ValentijnvdBeek
force-pushed
the
vvandebe.shufflevector.pattern.optimization
branch
from
73a92c2 to
e160d1c
Compare
ValentijnvdBeek
force-pushed
the
vvandebe.shufflevector.pattern.optimization
branch
from
e160d1c to
07244e7
Compare
ValentijnvdBeek
force-pushed
the
vvandebe.shufflevector.pattern.optimization
branch
from
07244e7 to
4836f6a
Compare
ValentijnvdBeek
force-pushed
the
vvandebe.vshuffle.impl
branch
from
5654047 to
84f3995
Compare
ValentijnvdBeek
force-pushed
the
vvandebe.shufflevector.pattern.optimization
branch
2 times, most recently
from
b41f4e1 to
ea44d18
Compare
ValentijnvdBeek
force-pushed
the
vvandebe.vshuffle.impl
branch
from
84f3995 to
5c3b1a6
Compare
ValentijnvdBeek
force-pushed
the
vvandebe.shufflevector.pattern.optimization
branch
2 times, most recently
from
4c022df to
b49d34c
Compare
ValentijnvdBeek
force-pushed
the
vvandebe.shufflevector.pattern.optimization
branch
2 times, most recently
from
f855c29 to
4d6af83
Compare
We check for iterative shift masks which corresponds to the CONCAT_VECTOR instruction.
…hunks of a vector
… of two vectors together
ValentijnvdBeek
force-pushed
the
vvandebe.shufflevector.pattern.optimization
branch
from
4d6af83 to
13cc82a
Compare
ValentijnvdBeek
force-pushed
the
vvandebe.vshuffle.impl
branch
2 times, most recently
from
e05b018 to
aec1600
Compare
ValentijnvdBeek
force-pushed
the
vvandebe.vshuffle.impl
branch
from
aec1600 to
d1d0a3a
Compare
ValentijnvdBeek
commented
Aug 15, 2024
Comment on lines
450
to
461
| --- | ||
| name: concat_vector_reverse_32_512_random | ||
| legalized: false | ||
| body: | | ||
| bb.1.entry: | ||
| liveins: $wl2, $wl4 | ||
| ; CHECK-LABEL: name: concat_vector_reverse_32_512_random | ||
| ; CHECK: liveins: $wl2, $wl4 | ||
| ; CHECK-NEXT: {{ $}} | ||
| ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(<8 x s32>) = COPY $wl2 | ||
| ; CHECK-NEXT: [[COPY1:%[0-9]+]]:_(<8 x s32>) = COPY $wl4 | ||
| ; CHECK-NEXT: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s32>) = G_CONCAT_VECTORS [[COPY1]](<8 x s32>), [[COPY]](<8 x s32>) |
There was a problem hiding this comment.
These got moved somehow, I will fix it when i get to it
ValentijnvdBeek
force-pushed
the
vvandebe.shufflevector.pattern.optimization
branch
2 times, most recently
from
c38562a to
ebe6489
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This merge requests adds the same framework for Shuffle Vector combinations used in #41 to the AIE2 backend. It also defines a new generator that generates the pattern used for a decent for subset of the shuffle vector modes. Concretely it matches mode 35 (4x4 -> 4x4 matrix transpose) and 29 (8x4->4x8 matrix transpose). Finally, it adds a generic opcode for the AIE vshuffle instruction.