This is an archive of the discontinued LLVM Phabricator instance.

Differential D136286

[mlir][sparse] use straightline and loop to insert into tensor
ClosedPublic

Authored by aartbik on Oct 19 2022, 12:45 PM.

Details

Reviewers
nicolasvasilache
Peiming
Commits
rGb430a352eff3: [mlir][sparse] use straightline and loop to insert into tensor
Summary

This exposed a missing type conversion for codegen

Diff Detail

Event Timeline

aartbik created this revision.Oct 19 2022, 12:45 PM
Herald added a project: Restricted Project. · View Herald TranscriptOct 19 2022, 12:45 PM
Herald added subscribers: anlunx, bzcheeseman, sdasgup3 and 19 others. · View Herald Transcript
aartbik requested review of this revision.Oct 19 2022, 12:45 PM
Herald added a reviewer: nicolasvasilache. · View Herald TranscriptOct 19 2022, 12:45 PM
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: stephenneuendorffer, nicolasvasilache. · View Herald Transcript
aartbik updated this revision to Diff 469015.Oct 19 2022, 12:48 PM

fixed header

aartbik added a reviewer: Peiming.Oct 19 2022, 12:48 PM
Harbormaster completed remote builds in B193073: Diff 469015.Oct 19 2022, 1:04 PM
aartbik added inline comments.Oct 19 2022, 6:30 PM
mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_insert_1d.mlir
65

Just FYI at the moment. We break here because

(1) bufferization copy
(2) SCF.for does not allow 1:N type conversions yet

aartbik mentioned this in D136314: [mlir][scf] support 1:N type conversion for scf.for..Oct 19 2022, 9:07 PM
Peiming mentioned this in D136372: [mlir][sparse] attach bufferizableOpInterface to InsertOp.Oct 20 2022, 12:42 PM
Peiming added inline comments.Oct 20 2022, 1:05 PM
mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_insert_1d.mlir
70

You forget a sparse_tensor.load %7 here

aartbik marked an inline comment as done.Oct 20 2022, 1:08 PM
aartbik added inline comments.
mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_insert_1d.mlir
70

I sure did!

aartbik updated this revision to Diff 469333.Oct 20 2022, 1:22 PM
aartbik marked an inline comment as done.

load + hasInserts

Harbormaster completed remote builds in B193310: Diff 469333.Oct 20 2022, 1:37 PM
Peiming mentioned this in rGfd7469404bcd: [mlir][sparse] attach bufferizableOpInterface to InsertOp.Oct 20 2022, 1:54 PM
aartbik updated this revision to Diff 469684.Oct 21 2022, 10:46 AM

rebased with main, added NOE usage

(note that this of course still needs the 1:N conversion)

Harbormaster completed remote builds in B193577: Diff 469684.Oct 21 2022, 11:17 AM
aartbik updated this revision to Diff 469784.Oct 21 2022, 3:08 PM

rebased with main, functional now

Harbormaster completed remote builds in B193660: Diff 469784.Oct 21 2022, 3:31 PM
Peiming accepted this revision.Oct 21 2022, 4:09 PM
This revision is now accepted and ready to land.Oct 21 2022, 4:09 PM
This revision was landed with ongoing or failed builds.Oct 21 2022, 4:17 PM
Closed by commit rGb430a352eff3: [mlir][sparse] use straightline and loop to insert into tensor (authored by aartbik). · Explain Why
This revision was automatically updated to reflect the committed changes.
aartbik added a commit: rGb430a352eff3: [mlir][sparse] use straightline and loop to insert into tensor.

Revision Contents

PathSize
mlir/
test/
Integration/
Dialect/
SparseTensor/
CPU/
27 lines

Diff 469815

mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_insert_1d.mlir

Show All 38 Linesmodule {
} }
func.func @entry() { func.func @entry() {
%f1 = arith.constant 1.0 : f32 %f1 = arith.constant 1.0 : f32
%f2 = arith.constant 2.0 : f32 %f2 = arith.constant 2.0 : f32
%c0 = arith.constant 0 : index %c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index %c1 = arith.constant 1 : index
%c3 = arith.constant 3 : index %c3 = arith.constant 3 : index
%c8 = arith.constant 8 : index
%c1023 = arith.constant 1023 : index %c1023 = arith.constant 1023 : index
// Build the sparse vector from code. // Build the sparse vector from straightline code.
%0 = bufferization.alloc_tensor() : tensor<1024xf32, #SparseVector> %0 = bufferization.alloc_tensor() : tensor<1024xf32, #SparseVector>
%1 = sparse_tensor.insert %f1 into %0[%c0] : tensor<1024xf32, #SparseVector> %1 = sparse_tensor.insert %f1 into %0[%c0] : tensor<1024xf32, #SparseVector>
%2 = sparse_tensor.insert %f2 into %1[%c1] : tensor<1024xf32, #SparseVector> %2 = sparse_tensor.insert %f2 into %1[%c1] : tensor<1024xf32, #SparseVector>
%3 = sparse_tensor.insert %f1 into %2[%c3] : tensor<1024xf32, #SparseVector> %3 = sparse_tensor.insert %f1 into %2[%c3] : tensor<1024xf32, #SparseVector>
%4 = sparse_tensor.insert %f2 into %3[%c1023] : tensor<1024xf32, #SparseVector> %4 = sparse_tensor.insert %f2 into %3[%c1023] : tensor<1024xf32, #SparseVector>
%5 = sparse_tensor.load %4 hasInserts : tensor<1024xf32, #SparseVector> %5 = sparse_tensor.load %4 hasInserts : tensor<1024xf32, #SparseVector>
// CHECK: ( 0, 4, 99, 99, 99, 99, 99, 99 ) // CHECK: ( 0, 4, 99, 99, 99, 99, 99, 99 )
// CHECK-NEXT: ( 0, 1, 3, 1023, 99, 99, 99, 99 ) // CHECK-NEXT: ( 0, 1, 3, 1023, 99, 99, 99, 99 )
// CHECK-NEXT: ( 1, 2, 1, 2, 99, 99, 99, 99 ) // CHECK-NEXT: ( 1, 2, 1, 2, 99, 99, 99, 99 )
call @dump(%5) : (tensor<1024xf32, #SparseVector>) -> () call @dump(%5) : (tensor<1024xf32, #SparseVector>) -> ()
// Build another sparse vector in a loop.
%6 = bufferization.alloc_tensor() : tensor<1024xf32, #SparseVector>
%7 = scf.for %i = %c0 to %c8 step %c1 iter_args(%vin = %6) -> tensor<1024xf32, #SparseVector> {
aartbikAuthorUnsubmitted
Done
ReplyInline Actions

Just FYI at the moment. We break here because

(1) bufferization copy
(2) SCF.for does not allow 1:N type conversions yet

aartbik: Just FYI at the moment. We break here because (1) bufferization copy (2) SCF.for does not…
%ii = arith.muli %i, %c3 : index
%vout = sparse_tensor.insert %f1 into %vin[%ii] : tensor<1024xf32, #SparseVector>
scf.yield %vout : tensor<1024xf32, #SparseVector>
}
%8 = sparse_tensor.load %7 hasInserts : tensor<1024xf32, #SparseVector>
PeimingUnsubmitted
Done
ReplyInline Actions

You forget a sparse_tensor.load %7 here

Peiming: You forget a `sparse_tensor.load %7` here
aartbikAuthorUnsubmitted
Done
ReplyInline Actions

I sure did!

aartbik: I sure did!
// CHECK: ( 0, 8, 99, 99, 99, 99, 99, 99 )
// CHECK-NEXT: ( 0, 3, 6, 9, 12, 15, 18, 21 )
// CHECK-NEXT: ( 1, 1, 1, 1, 1, 1, 1, 1 )
//
call @dump(%8) : (tensor<1024xf32, #SparseVector>) -> ()
// CHECK-NEXT: 4
// CHECK-NEXT: 8
%noe1 = sparse_tensor.number_of_entries %5 : tensor<1024xf32, #SparseVector>
%noe2 = sparse_tensor.number_of_entries %8 : tensor<1024xf32, #SparseVector>
vector.print %noe1 : index
vector.print %noe2 : index
// Free resources.
bufferization.dealloc_tensor %5 : tensor<1024xf32, #SparseVector> bufferization.dealloc_tensor %5 : tensor<1024xf32, #SparseVector>
bufferization.dealloc_tensor %8 : tensor<1024xf32, #SparseVector>
return return
} }
} }