This is an archive of the discontinued LLVM Phabricator instance.

Differential D107003

Rorder mmt4d iteration domain
ClosedPublic

Authored by asaadaldien on Jul 28 2021, 1:56 PM.

Details

Reviewers
mravishankar
nicolasvasilache
rsuderman
Commits
rG8a0d6e839f0f: Rorder mmt4d iteration domain
Summary

Move tile iterators to outer most dim

Diff Detail

Event Timeline

asaadaldien created this revision.Jul 28 2021, 1:56 PM
Herald added a reviewer: mravishankar. · View Herald TranscriptJul 28 2021, 1:56 PM
Herald added subscribers: shabalin, dcaballe, cota and 18 others. · View Herald Transcript
asaadaldien requested review of this revision.Jul 28 2021, 1:56 PM
Herald added a reviewer: nicolasvasilache. · View Herald TranscriptJul 28 2021, 1:56 PM
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: stephenneuendorffer, nicolasvasilache. · View Herald Transcript
asaadaldien added a reviewer: rsuderman.Jul 28 2021, 1:57 PM
Harbormaster completed remote builds in B116798: Diff 362532.Jul 28 2021, 4:43 PM
rsuderman accepted this revision.Jul 28 2021, 7:27 PM
This revision is now accepted and ready to land.Jul 28 2021, 7:27 PM
rsuderman added a comment.Jul 28 2021, 8:24 PM

Note typo in the comment header s/rorder/reorder

asaadaldien updated this revision to Diff 362814.Jul 29 2021, 9:58 AM

Rebase and fix commit msg

Herald added a subscriber: Chia-hungDuan. · View Herald TranscriptJul 29 2021, 9:58 AM
Harbormaster completed remote builds in B117000: Diff 362814.Jul 29 2021, 10:43 AM
Closed by commit rG8a0d6e839f0f: Rorder mmt4d iteration domain (authored by asaadaldien). · Explain WhyJul 29 2021, 11:32 AM
This revision was automatically updated to reflect the committed changes.
asaadaldien added a commit: rG8a0d6e839f0f: Rorder mmt4d iteration domain.
Benoit mentioned this in D109692: Reorder mmt4d shapes:.Sep 13 2021, 8:11 AM
GMNGeoffrey mentioned this in rG340314c4dcc8: Reorder mmt4d shapes:.Sep 13 2021, 12:10 PM

Revision Contents

PathSize
mlir/
include/
mlir/
Dialect/
Linalg/
IR/
7 lines
python/
mlir/
dialects/
linalg/
opdsl/
ops/
2 lines

Diff 362833

mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml

Show First 20 Lines • Show All 182 Lines • ▼ Show 20 Lines - !LinalgOperandDefConfig
shape_map: affine_map<()[s0, s1, s2, s3, s4, s5] -> (s4, s1, s5, s3)> shape_map: affine_map<()[s0, s1, s2, s3, s4, s5] -> (s4, s1, s5, s3)>
- !LinalgOperandDefConfig - !LinalgOperandDefConfig
name: accum name: accum
usage: OutputOperand usage: OutputOperand
type_var: AccumType type_var: AccumType
shape_map: affine_map<()[s0, s1, s2, s3, s4, s5] -> (s0, s4, s2, s5)> shape_map: affine_map<()[s0, s1, s2, s3, s4, s5] -> (s0, s4, s2, s5)>
indexing_maps: !LinalgIndexingMapsConfig indexing_maps: !LinalgIndexingMapsConfig
static_indexing_maps: static_indexing_maps:
- affine_map<(d0, d1, d2, d3, d4, d5)[s0, s1, s2, s3, s4, s5] -> (d0, d4, d1, - affine_map<(d0, d1, d2, d3, d4, d5)[s0, s1, s2, s3, s4, s5] -> (d0, d4, d2,
d5)> d5)>
- affine_map<(d0, d1, d2, d3, d4, d5)[s0, s1, s2, s3, s4, s5] -> (d2, d4, d3, - affine_map<(d0, d1, d2, d3, d4, d5)[s0, s1, s2, s3, s4, s5] -> (d1, d4, d3,
d5)> d5)>
- affine_map<(d0, d1, d2, d3, d4, d5)[s0, s1, s2, s3, s4, s5] -> (d0, d2, d1, - affine_map<(d0, d1, d2, d3, d4, d5)[s0, s1, s2, s3, s4, s5] -> (d0, d1, d2,
d3)> d3)>
iterator_types: iterator_types:
- parallel - parallel
- parallel - parallel
- parallel - parallel
- parallel - parallel
- reduction - reduction
- reduction - reduction
▲ Show 20 Lines • Show All 1,491 Lines • ▼ Show 20 Lines value: !ScalarExpression
fn_name: exp fn_name: exp
operands: operands:
- !ScalarExpression - !ScalarExpression
symbolic_cast: symbolic_cast:
type_var: U type_var: U
operands: operands:
- !ScalarExpression - !ScalarExpression
scalar_arg: I scalar_arg: I

mlir/python/mlir/dialects/linalg/opdsl/ops/core_named_ops.py

Show First 20 Lines • Show All 46 Lines • ▼ Show 20 Lines """Performs a matrix-matrix-transpose multiplication of two 4D inputs.
Differences from linalg.matmul: Differences from linalg.matmul:
* The right hand side is transposed, whence the 't' in 'mmt'. * The right hand side is transposed, whence the 't' in 'mmt'.
* The input and output tensors have a 4D shape instead of a 2D shape. They * The input and output tensors have a 4D shape instead of a 2D shape. They
are interpreted as 2D matrices with one level of 2D tile subdivision, are interpreted as 2D matrices with one level of 2D tile subdivision,
whence the 2+2=4 dimensions. The inner tile dimensions are identified with whence the 2+2=4 dimensions. The inner tile dimensions are identified with
'0' suffixes below, for instance the LHS matrix shape (M, K, M0, K0) reads '0' suffixes below, for instance the LHS matrix shape (M, K, M0, K0) reads
as: MxK tiles, each of shape M0xK0. as: MxK tiles, each of shape M0xK0.
""" """
domain(D.m, D.m0, D.n, D.n0, D.k, D.k0) domain(D.m, D.n, D.m0, D.n0, D.k, D.k0)
implements(ContractionOpInterface) implements(ContractionOpInterface)
accum[D.m, D.n, D.m0, D.n0] += cast(TV.AccumType, lhs[D.m, D.k, D.m0, D.k0]) * cast(TV.AccumType, rhs[D.n, D.k, D.n0, D.k0]) accum[D.m, D.n, D.m0, D.n0] += cast(TV.AccumType, lhs[D.m, D.k, D.m0, D.k0]) * cast(TV.AccumType, rhs[D.n, D.k, D.n0, D.k0])
@linalg_structured_op @linalg_structured_op
def batch_matmul( def batch_matmul(
A=TensorDef(T1, Batch, S.M, S.K), A=TensorDef(T1, Batch, S.M, S.K),
B=TensorDef(T2, Batch, S.K, S.N), B=TensorDef(T2, Batch, S.K, S.N),
C=TensorDef(U, Batch, S.M, S.N, output=True)): C=TensorDef(U, Batch, S.M, S.N, output=True)):
▲ Show 20 Lines • Show All 325 Lines • Show Last 20 Lines