This is an archive of the discontinued LLVM Phabricator instance.

Differential D139625

[mlir][SparseTensor] Add a few tests for sparse vectorization
ClosedPublic

Authored by qcolombet on Dec 8 2022, 5:08 AM.

Details

Reviewers
aartbik
wrengr
Peiming
bixia
nicolasvasilache
Commits
rGa0220ba7218f: [mlir][SparseTensor] Add a few tests for sparse vectorization
Summary

These tests covers mulf, ori, and subi.

NFC

Note: I plan to add more tests to cover the other operators as well (div, add, xor, etc.) but I wanted to double check that the direction was okay first.

BTW, I may have found a bug in the sparsification pass. I found a case where it crashes, but maybe it's because my hand-generated IR is somehow invalid. See https://github.com/llvm/llvm-project/issues/59394 for more details.

Diff Detail

Event Timeline

qcolombet created this revision.Dec 8 2022, 5:08 AM
Herald added a project: Restricted Project. · View Herald TranscriptDec 8 2022, 5:08 AM
Herald added subscribers: hanchung, jsetoain, Moerafaat and 21 others. · View Herald Transcript
qcolombet requested review of this revision.Dec 8 2022, 5:08 AM
Herald added a reviewer: nicolasvasilache. · View Herald TranscriptDec 8 2022, 5:08 AM
Herald added subscribers: • pcwang-thead, stephenneuendorffer, nicolasvasilache. · View Herald Transcript
aartbik added inline comments.Dec 8 2022, 10:16 AM
mlir/test/Dialect/SparseTensor/vectorize_reduction.mlir
3

can we make this CHECK-ON/CHECK-OFF to make it more consistent with style in this dir?

364

wer -> we

Harbormaster completed remote builds in B201944: Diff 481251.Dec 8 2022, 12:07 PM
qcolombet updated this revision to Diff 481619.Dec 9 2022, 6:16 AM
  • Prepend CHECK- to check prefixes.
qcolombet marked 2 inline comments as done.Dec 9 2022, 6:16 AM
qcolombet added inline comments.
mlir/test/Dialect/SparseTensor/vectorize_reduction.mlir
3

Sure thing!

364

Good catch

Harbormaster completed remote builds in B202214: Diff 481619.Dec 9 2022, 6:36 AM
aartbik added inline comments.Dec 9 2022, 4:06 PM
mlir/test/Dialect/SparseTensor/vectorize_reduction.mlir
439

can you remove this part from the test now?

qcolombet marked 2 inline comments as done.Dec 9 2022, 4:17 PM
qcolombet added inline comments.
mlir/test/Dialect/SparseTensor/vectorize_reduction.mlir
439

Sure thing!
I’ll do that next week

qcolombet updated this revision to Diff 482016.Dec 12 2022, 12:49 AM
  • Remove commented test
qcolombet marked an inline comment as done.Dec 12 2022, 12:49 AM
Harbormaster completed remote builds in B202503: Diff 482016.Dec 12 2022, 1:17 AM
aartbik accepted this revision.Dec 14 2022, 9:11 AM
This revision is now accepted and ready to land.Dec 14 2022, 9:11 AM
Closed by commit rGa0220ba7218f: [mlir][SparseTensor] Add a few tests for sparse vectorization (authored by qcolombet). · Explain WhyDec 14 2022, 11:05 AM
This revision was automatically updated to reflect the committed changes.
qcolombet added a commit: rGa0220ba7218f: [mlir][SparseTensor] Add a few tests for sparse vectorization.

Revision Contents

PathSize
mlir/
test/
Dialect/
SparseTensor/
436 lines

Diff 482938

mlir/test/Dialect/SparseTensor/vectorize_reduction.mlir

  • This file was added.
// RUN: mlir-opt %s -sparsification -cse -sparse-vectorization="vl=8" -cse -split-input-file | \
// RUN: FileCheck %s --check-prefix=CHECK-ON
// RUN: mlir-opt %s -sparsification -cse -split-input-file | \
aartbikUnsubmitted
Done
ReplyInline Actions

can we make this CHECK-ON/CHECK-OFF to make it more consistent with style in this dir?

aartbik: can we make this CHECK-ON/CHECK-OFF to make it more consistent with style in this dir?
qcolombetAuthorUnsubmitted
Done
ReplyInline Actions

Sure thing!

qcolombet: Sure thing!
// RUN: FileCheck %s --check-prefix=CHECK-OFF
// -----
// Check that we recognize a reduction with a mul operator.
// We use two dimensions here to check that the vectorization
// is not affected by how the outer loop is layed out.
// In other words, we should be able to vectorize the sparse inner loop
// regardless of whether the outer loop is dense or sparse.
//
// For this particular test, we expect:
// With vectorization on:
// dense scf.for
// init vector_accumulator = {scalar_accumulator, 1.0, 1.0, ...}
// sparse scf.for
// vectorized mul in vector_accumulator, vector_input
// horizontal reduction of the vector_accumulator to scalar_accumulator
// final store of scalar_accumulaor
//
// With vectorization off:
// dense scf.for
// sparse scf.for
// mul in accumulator
// final store
//
// CHECK-ON-LABEL: func.func @sparse_product_reduction_dense_sparse(
// CHECK-ON-SAME: %[[VAL_0:.*]]: tensor<f64>,
// CHECK-ON-SAME: %[[VAL_1:.*]]: tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>>) -> tensor<f64> {
// CHECK-ON-DAG: %[[VAL_2:.*]] = arith.constant 8 : index
// CHECK-ON-DAG: %[[VAL_3:.*]] = arith.constant dense<1.000000e+00> : vector<8xf64>
// CHECK-ON-DAG: %[[VAL_4:.*]] = arith.constant dense<0.000000e+00> : vector<8xf64>
// CHECK-ON-DAG: %[[VAL_5:.*]] = arith.constant 0 : index
// CHECK-ON-DAG: %[[VAL_6:.*]] = arith.constant 1 : index
// CHECK-ON-DAG: %[[VAL_7:.*]] = tensor.dim %[[VAL_1]], %[[VAL_5]] : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>>
// CHECK-ON: %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 1 : index} : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xindex>
// CHECK-ON: %[[VAL_9:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xf64>
// CHECK-ON: %[[VAL_10:.*]] = bufferization.to_memref %[[VAL_0]] : memref<f64>
// CHECK-ON: %[[VAL_11:.*]] = memref.load %[[VAL_10]][] : memref<f64>
// CHECK-ON: %[[VAL_12:.*]] = scf.for %[[VAL_13:.*]] = %[[VAL_5]] to %[[VAL_7]] step %[[VAL_6]] iter_args(%[[VAL_14:.*]] = %[[VAL_11]]) -> (f64) {
// CHECK-ON: %[[VAL_15:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_13]]] : memref<?xindex>
// CHECK-ON: %[[VAL_16:.*]] = arith.addi %[[VAL_13]], %[[VAL_6]] : index
// CHECK-ON: %[[VAL_17:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_16]]] : memref<?xindex>
// CHECK-ON: %[[VAL_18:.*]] = vector.insertelement %[[VAL_14]], %[[VAL_3]]{{\[}}%[[VAL_5]] : index] : vector<8xf64>
// CHECK-ON: %[[VAL_19:.*]] = scf.for %[[VAL_20:.*]] = %[[VAL_15]] to %[[VAL_17]] step %[[VAL_2]] iter_args(%[[VAL_21:.*]] = %[[VAL_18]]) -> (vector<8xf64>) {
// CHECK-ON: %[[VAL_22:.*]] = affine.min #map(%[[VAL_17]], %[[VAL_20]]){{\[}}%[[VAL_2]]]
// CHECK-ON: %[[VAL_23:.*]] = vector.create_mask %[[VAL_22]] : vector<8xi1>
// CHECK-ON: %[[VAL_24:.*]] = vector.maskedload %[[VAL_9]]{{\[}}%[[VAL_20]]], %[[VAL_23]], %[[VAL_4]] : memref<?xf64>, vector<8xi1>, vector<8xf64> into vector<8xf64>
// CHECK-ON: %[[VAL_25:.*]] = arith.mulf %[[VAL_21]], %[[VAL_24]] : vector<8xf64>
// CHECK-ON: %[[VAL_26:.*]] = arith.select %[[VAL_23]], %[[VAL_25]], %[[VAL_21]] : vector<8xi1>, vector<8xf64>
// CHECK-ON: scf.yield %[[VAL_26]] : vector<8xf64>
// CHECK-ON: } {"Emitted from" = "linalg.generic"}
// CHECK-ON: %[[VAL_27:.*]] = vector.reduction <mul>, %[[VAL_28:.*]] : vector<8xf64> into f64
// CHECK-ON: scf.yield %[[VAL_27]] : f64
// CHECK-ON: } {"Emitted from" = "linalg.generic"}
// CHECK-ON: memref.store %[[VAL_29:.*]], %[[VAL_10]][] : memref<f64>
// CHECK-ON: %[[VAL_30:.*]] = bufferization.to_tensor %[[VAL_10]] : memref<f64>
// CHECK-ON: return %[[VAL_30]] : tensor<f64>
// CHECK-ON: }
//
// CHECK-OFF-LABEL: func.func @sparse_product_reduction_dense_sparse(
// CHECK-OFF-SAME: %[[VAL_0:.*]]: tensor<f64>,
// CHECK-OFF-SAME: %[[VAL_1:.*]]: tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>>) -> tensor<f64> {
// CHECK-OFF-DAG: %[[VAL_2:.*]] = arith.constant 0 : index
// CHECK-OFF-DAG: %[[VAL_3:.*]] = arith.constant 1 : index
// CHECK-OFF: %[[VAL_4:.*]] = tensor.dim %[[VAL_1]], %[[VAL_2]] : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>>
// CHECK-OFF: %[[VAL_5:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 1 : index} : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xindex>
// CHECK-OFF: %[[VAL_6:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ] }>> to memref<?xf64>
// CHECK-OFF: %[[VAL_7:.*]] = bufferization.to_memref %[[VAL_0]] : memref<f64>
// CHECK-OFF: %[[VAL_8:.*]] = memref.load %[[VAL_7]][] : memref<f64>
// CHECK-OFF: %[[VAL_9:.*]] = scf.for %[[VAL_10:.*]] = %[[VAL_2]] to %[[VAL_4]] step %[[VAL_3]] iter_args(%[[VAL_11:.*]] = %[[VAL_8]]) -> (f64) {
// CHECK-OFF: %[[VAL_12:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_10]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_13:.*]] = arith.addi %[[VAL_10]], %[[VAL_3]] : index
// CHECK-OFF: %[[VAL_14:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_13]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_15:.*]] = scf.for %[[VAL_16:.*]] = %[[VAL_12]] to %[[VAL_14]] step %[[VAL_3]] iter_args(%[[VAL_17:.*]] = %[[VAL_11]]) -> (f64) {
// CHECK-OFF: %[[VAL_18:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_16]]] : memref<?xf64>
// CHECK-OFF: %[[VAL_19:.*]] = arith.mulf %[[VAL_17]], %[[VAL_18]] : f64
// CHECK-OFF: scf.yield %[[VAL_19]] : f64
// CHECK-OFF: } {"Emitted from" = "linalg.generic"}
// CHECK-OFF: scf.yield %[[VAL_20:.*]] : f64
// CHECK-OFF: } {"Emitted from" = "linalg.generic"}
// CHECK-OFF: memref.store %[[VAL_21:.*]], %[[VAL_7]][] : memref<f64>
// CHECK-OFF: %[[VAL_22:.*]] = bufferization.to_tensor %[[VAL_7]] : memref<f64>
// CHECK-OFF: return %[[VAL_22]] : tensor<f64>
// CHECK-OFF: }
#SparseVector = #sparse_tensor.encoding<{dimLevelType = ["dense","compressed"]}>
#trait = {
indexing_maps = [
affine_map<(i,j) -> (i,j)>, // a (in)
affine_map<(i,j) -> ()> // x (out)
],
iterator_types = ["reduction", "reduction"]
}
func.func @sparse_product_reduction_dense_sparse(%argx: tensor<f64>,
%arga: tensor<?x128xf64, #SparseVector>)
-> tensor<f64> {
%0 = linalg.generic #trait
ins(%arga: tensor<?x128xf64, #SparseVector>)
outs(%argx: tensor<f64>) {
^bb(%a: f64, %x: f64):
%t = arith.mulf %x, %a: f64
linalg.yield %t : f64
} -> tensor<f64>
return %0 : tensor<f64>
}
// -----
// Same as sparse_product_reduction_dense_sparse but with the outer loop being sparse.
//
// CHECK-ON-LABEL: func.func @sparse_product_reduction_sparse_sparse(
// CHECK-ON-SAME: %[[VAL_0:.*]]: tensor<f64>,
// CHECK-ON-SAME: %[[VAL_1:.*]]: tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>>) -> tensor<f64> {
// CHECK-ON-DAG: %[[VAL_2:.*]] = arith.constant 8 : index
// CHECK-ON-DAG: %[[VAL_3:.*]] = arith.constant dense<1.000000e+00> : vector<8xf64>
// CHECK-ON-DAG: %[[VAL_4:.*]] = arith.constant dense<0.000000e+00> : vector<8xf64>
// CHECK-ON-DAG: %[[VAL_5:.*]] = arith.constant 0 : index
// CHECK-ON-DAG: %[[VAL_6:.*]] = arith.constant 1 : index
// CHECK-ON: %[[VAL_7:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>> to memref<?xindex>
// CHECK-ON: %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 1 : index} : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>> to memref<?xindex>
// CHECK-ON: %[[VAL_9:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>> to memref<?xf64>
// CHECK-ON: %[[VAL_10:.*]] = bufferization.to_memref %[[VAL_0]] : memref<f64>
// CHECK-ON: %[[VAL_11:.*]] = memref.load %[[VAL_10]][] : memref<f64>
// CHECK-ON: %[[VAL_12:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_5]]] : memref<?xindex>
// CHECK-ON: %[[VAL_13:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_6]]] : memref<?xindex>
// CHECK-ON: %[[VAL_14:.*]] = scf.for %[[VAL_15:.*]] = %[[VAL_12]] to %[[VAL_13]] step %[[VAL_6]] iter_args(%[[VAL_16:.*]] = %[[VAL_11]]) -> (f64) {
// CHECK-ON: %[[VAL_17:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_15]]] : memref<?xindex>
// CHECK-ON: %[[VAL_18:.*]] = arith.addi %[[VAL_15]], %[[VAL_6]] : index
// CHECK-ON: %[[VAL_19:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_18]]] : memref<?xindex>
// CHECK-ON: %[[VAL_20:.*]] = vector.insertelement %[[VAL_16]], %[[VAL_3]]{{\[}}%[[VAL_5]] : index] : vector<8xf64>
// CHECK-ON: %[[VAL_21:.*]] = scf.for %[[VAL_22:.*]] = %[[VAL_17]] to %[[VAL_19]] step %[[VAL_2]] iter_args(%[[VAL_23:.*]] = %[[VAL_20]]) -> (vector<8xf64>) {
// CHECK-ON: %[[VAL_24:.*]] = affine.min #map(%[[VAL_19]], %[[VAL_22]]){{\[}}%[[VAL_2]]]
// CHECK-ON: %[[VAL_25:.*]] = vector.create_mask %[[VAL_24]] : vector<8xi1>
// CHECK-ON: %[[VAL_26:.*]] = vector.maskedload %[[VAL_9]]{{\[}}%[[VAL_22]]], %[[VAL_25]], %[[VAL_4]] : memref<?xf64>, vector<8xi1>, vector<8xf64> into vector<8xf64>
// CHECK-ON: %[[VAL_27:.*]] = arith.mulf %[[VAL_23]], %[[VAL_26]] : vector<8xf64>
// CHECK-ON: %[[VAL_28:.*]] = arith.select %[[VAL_25]], %[[VAL_27]], %[[VAL_23]] : vector<8xi1>, vector<8xf64>
// CHECK-ON: scf.yield %[[VAL_28]] : vector<8xf64>
// CHECK-ON: } {"Emitted from" = "linalg.generic"}
// CHECK-ON: %[[VAL_29:.*]] = vector.reduction <mul>, %[[VAL_30:.*]] : vector<8xf64> into f64
// CHECK-ON: scf.yield %[[VAL_29]] : f64
// CHECK-ON: } {"Emitted from" = "linalg.generic"}
// CHECK-ON: memref.store %[[VAL_31:.*]], %[[VAL_10]][] : memref<f64>
// CHECK-ON: %[[VAL_32:.*]] = bufferization.to_tensor %[[VAL_10]] : memref<f64>
// CHECK-ON: return %[[VAL_32]] : tensor<f64>
// CHECK-ON: }
//
// CHECK-OFF-LABEL: func.func @sparse_product_reduction_sparse_sparse(
// CHECK-OFF-SAME: %[[VAL_0:.*]]: tensor<f64>,
// CHECK-OFF-SAME: %[[VAL_1:.*]]: tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>>) -> tensor<f64> {
// CHECK-OFF-DAG: %[[VAL_2:.*]] = arith.constant 0 : index
// CHECK-OFF-DAG: %[[VAL_3:.*]] = arith.constant 1 : index
// CHECK-OFF: %[[VAL_4:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>> to memref<?xindex>
// CHECK-OFF: %[[VAL_5:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 1 : index} : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>> to memref<?xindex>
// CHECK-OFF: %[[VAL_6:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?x128xf64, #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>> to memref<?xf64>
// CHECK-OFF: %[[VAL_7:.*]] = bufferization.to_memref %[[VAL_0]] : memref<f64>
// CHECK-OFF: %[[VAL_8:.*]] = memref.load %[[VAL_7]][] : memref<f64>
// CHECK-OFF: %[[VAL_9:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_2]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_10:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_3]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_11:.*]] = scf.for %[[VAL_12:.*]] = %[[VAL_9]] to %[[VAL_10]] step %[[VAL_3]] iter_args(%[[VAL_13:.*]] = %[[VAL_8]]) -> (f64) {
// CHECK-OFF: %[[VAL_14:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_12]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_15:.*]] = arith.addi %[[VAL_12]], %[[VAL_3]] : index
// CHECK-OFF: %[[VAL_16:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_15]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_17:.*]] = scf.for %[[VAL_18:.*]] = %[[VAL_14]] to %[[VAL_16]] step %[[VAL_3]] iter_args(%[[VAL_19:.*]] = %[[VAL_13]]) -> (f64) {
// CHECK-OFF: %[[VAL_20:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_18]]] : memref<?xf64>
// CHECK-OFF: %[[VAL_21:.*]] = arith.mulf %[[VAL_19]], %[[VAL_20]] : f64
// CHECK-OFF: scf.yield %[[VAL_21]] : f64
// CHECK-OFF: } {"Emitted from" = "linalg.generic"}
// CHECK-OFF: scf.yield %[[VAL_22:.*]] : f64
// CHECK-OFF: } {"Emitted from" = "linalg.generic"}
// CHECK-OFF: memref.store %[[VAL_23:.*]], %[[VAL_7]][] : memref<f64>
// CHECK-OFF: %[[VAL_24:.*]] = bufferization.to_tensor %[[VAL_7]] : memref<f64>
// CHECK-OFF: return %[[VAL_24]] : tensor<f64>
// CHECK-OFF: }
#SparseVector = #sparse_tensor.encoding<{dimLevelType = ["compressed","compressed"]}>
#trait = {
indexing_maps = [
affine_map<(i,j) -> (i,j)>, // a (in)
affine_map<(i,j) -> ()> // x (out)
],
iterator_types = ["reduction", "reduction"]
}
func.func @sparse_product_reduction_sparse_sparse(%argx: tensor<f64>,
%arga: tensor<?x128xf64, #SparseVector>)
-> tensor<f64> {
%0 = linalg.generic #trait
ins(%arga: tensor<?x128xf64, #SparseVector>)
outs(%argx: tensor<f64>) {
^bb(%a: f64, %x: f64):
%t = arith.mulf %x, %a: f64
linalg.yield %t : f64
} -> tensor<f64>
return %0 : tensor<f64>
}
// -----
// sparse_product_reduction_dense_sparse and
// sparse_product_reduction_sparse_sparse established that the outer loop
// doesn't matter for vectorization.
// As a result from this point forward, use tensors with fewer dimensions.
// Check that we vectorize reductions with ori.
// Note: The weird element type here is to check that we create the right
// constant type for the pass-through value.
// CHECK-ON-LABEL: func.func @sparse_reduction_ori(
// CHECK-ON-SAME: %[[VAL_0:.*]]: tensor<i13>,
// CHECK-ON-SAME: %[[VAL_1:.*]]: tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>>) -> tensor<i13> {
// CHECK-ON-DAG: %[[VAL_2:.*]] = arith.constant 8 : index
// CHECK-ON-DAG: %[[VAL_3:.*]] = arith.constant dense<0> : vector<8xi13>
// CHECK-ON-DAG: %[[VAL_4:.*]] = arith.constant 0 : index
// CHECK-ON-DAG: %[[VAL_5:.*]] = arith.constant 1 : index
// CHECK-ON: %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
// CHECK-ON: %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xi13>
// CHECK-ON: %[[VAL_8:.*]] = bufferization.to_memref %[[VAL_0]] : memref<i13>
// CHECK-ON: %[[VAL_9:.*]] = memref.load %[[VAL_8]][] : memref<i13>
// CHECK-ON: %[[VAL_10:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
// CHECK-ON: %[[VAL_11:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
// CHECK-ON: %[[VAL_12:.*]] = vector.broadcast %[[VAL_9]] : i13 to vector<8xi13>
// CHECK-ON: %[[VAL_13:.*]] = scf.for %[[VAL_14:.*]] = %[[VAL_10]] to %[[VAL_11]] step %[[VAL_2]] iter_args(%[[VAL_15:.*]] = %[[VAL_12]]) -> (vector<8xi13>) {
// CHECK-ON: %[[VAL_16:.*]] = affine.min #map(%[[VAL_11]], %[[VAL_14]]){{\[}}%[[VAL_2]]]
// CHECK-ON: %[[VAL_17:.*]] = vector.create_mask %[[VAL_16]] : vector<8xi1>
// CHECK-ON: %[[VAL_18:.*]] = vector.maskedload %[[VAL_7]]{{\[}}%[[VAL_14]]], %[[VAL_17]], %[[VAL_3]] : memref<?xi13>, vector<8xi1>, vector<8xi13> into vector<8xi13>
// CHECK-ON: %[[VAL_19:.*]] = arith.ori %[[VAL_15]], %[[VAL_18]] : vector<8xi13>
// CHECK-ON: %[[VAL_20:.*]] = arith.select %[[VAL_17]], %[[VAL_19]], %[[VAL_15]] : vector<8xi1>, vector<8xi13>
// CHECK-ON: scf.yield %[[VAL_20]] : vector<8xi13>
// CHECK-ON: } {"Emitted from" = "linalg.generic"}
// CHECK-ON: %[[VAL_21:.*]] = vector.reduction <or>, %[[VAL_22:.*]] : vector<8xi13> into i13
// CHECK-ON: memref.store %[[VAL_21]], %[[VAL_8]][] : memref<i13>
// CHECK-ON: %[[VAL_23:.*]] = bufferization.to_tensor %[[VAL_8]] : memref<i13>
// CHECK-ON: return %[[VAL_23]] : tensor<i13>
// CHECK-ON: }
//
// CHECK-OFF-LABEL: func.func @sparse_reduction_ori(
// CHECK-OFF-SAME: %[[VAL_0:.*]]: tensor<i13>,
// CHECK-OFF-SAME: %[[VAL_1:.*]]: tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>>) -> tensor<i13> {
// CHECK-OFF-DAG: %[[VAL_2:.*]] = arith.constant 0 : index
// CHECK-OFF-DAG: %[[VAL_3:.*]] = arith.constant 1 : index
// CHECK-OFF: %[[VAL_4:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
// CHECK-OFF: %[[VAL_5:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xi13>
// CHECK-OFF: %[[VAL_6:.*]] = bufferization.to_memref %[[VAL_0]] : memref<i13>
// CHECK-OFF: %[[VAL_7:.*]] = memref.load %[[VAL_6]][] : memref<i13>
// CHECK-OFF: %[[VAL_8:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_2]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_9:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_3]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_10:.*]] = scf.for %[[VAL_11:.*]] = %[[VAL_8]] to %[[VAL_9]] step %[[VAL_3]] iter_args(%[[VAL_12:.*]] = %[[VAL_7]]) -> (i13) {
// CHECK-OFF: %[[VAL_13:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_11]]] : memref<?xi13>
// CHECK-OFF: %[[VAL_14:.*]] = arith.ori %[[VAL_12]], %[[VAL_13]] : i13
// CHECK-OFF: scf.yield %[[VAL_14]] : i13
// CHECK-OFF: } {"Emitted from" = "linalg.generic"}
// CHECK-OFF: memref.store %[[VAL_15:.*]], %[[VAL_6]][] : memref<i13>
// CHECK-OFF: %[[VAL_16:.*]] = bufferization.to_tensor %[[VAL_6]] : memref<i13>
// CHECK-OFF: return %[[VAL_16]] : tensor<i13>
// CHECK-OFF: }
#SparseVector = #sparse_tensor.encoding<{dimLevelType = ["compressed"]}>
#trait = {
indexing_maps = [
affine_map<(i) -> (i)>, // a (in)
affine_map<(i) -> ()> // x (out)
],
iterator_types = ["reduction"]
}
func.func @sparse_reduction_ori(%argx: tensor<i13>,
%arga: tensor<?xi13, #SparseVector>)
-> tensor<i13> {
%0 = linalg.generic #trait
ins(%arga: tensor<?xi13, #SparseVector>)
outs(%argx: tensor<i13>) {
^bb(%a: i13, %x: i13):
%t = arith.ori %x, %a: i13
linalg.yield %t : i13
} -> tensor<i13>
return %0 : tensor<i13>
}
// -----
// Same test as sparse_reduction_ori except that the accumulator is on the
// rhs of the operation.
// This checks that we can recognize a reduction irrespective to where the
// accumalator appears on commutative operations.
// CHECK-ON-LABEL: func.func @sparse_reduction_ori_accumulator_on_rhs(
// CHECK-ON-SAME: %[[VAL_0:.*]]: tensor<i13>,
// CHECK-ON-SAME: %[[VAL_1:.*]]: tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>>) -> tensor<i13> {
// CHECK-ON-DAG: %[[VAL_2:.*]] = arith.constant 8 : index
// CHECK-ON-DAG: %[[VAL_3:.*]] = arith.constant dense<0> : vector<8xi13>
// CHECK-ON-DAG: %[[VAL_4:.*]] = arith.constant 0 : index
// CHECK-ON-DAG: %[[VAL_5:.*]] = arith.constant 1 : index
// CHECK-ON: %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
// CHECK-ON: %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xi13>
// CHECK-ON: %[[VAL_8:.*]] = bufferization.to_memref %[[VAL_0]] : memref<i13>
// CHECK-ON: %[[VAL_9:.*]] = memref.load %[[VAL_8]][] : memref<i13>
// CHECK-ON: %[[VAL_10:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex>
// CHECK-ON: %[[VAL_11:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
// CHECK-ON: %[[VAL_12:.*]] = vector.broadcast %[[VAL_9]] : i13 to vector<8xi13>
// CHECK-ON: %[[VAL_13:.*]] = scf.for %[[VAL_14:.*]] = %[[VAL_10]] to %[[VAL_11]] step %[[VAL_2]] iter_args(%[[VAL_15:.*]] = %[[VAL_12]]) -> (vector<8xi13>) {
// CHECK-ON: %[[VAL_16:.*]] = affine.min #map(%[[VAL_11]], %[[VAL_14]]){{\[}}%[[VAL_2]]]
// CHECK-ON: %[[VAL_17:.*]] = vector.create_mask %[[VAL_16]] : vector<8xi1>
// CHECK-ON: %[[VAL_18:.*]] = vector.maskedload %[[VAL_7]]{{\[}}%[[VAL_14]]], %[[VAL_17]], %[[VAL_3]] : memref<?xi13>, vector<8xi1>, vector<8xi13> into vector<8xi13>
// CHECK-ON: %[[VAL_19:.*]] = arith.ori %[[VAL_18]], %[[VAL_15]] : vector<8xi13>
// CHECK-ON: %[[VAL_20:.*]] = arith.select %[[VAL_17]], %[[VAL_19]], %[[VAL_15]] : vector<8xi1>, vector<8xi13>
// CHECK-ON: scf.yield %[[VAL_20]] : vector<8xi13>
// CHECK-ON: } {"Emitted from" = "linalg.generic"}
// CHECK-ON: %[[VAL_21:.*]] = vector.reduction <or>, %[[VAL_22:.*]] : vector<8xi13> into i13
// CHECK-ON: memref.store %[[VAL_21]], %[[VAL_8]][] : memref<i13>
// CHECK-ON: %[[VAL_23:.*]] = bufferization.to_tensor %[[VAL_8]] : memref<i13>
// CHECK-ON: return %[[VAL_23]] : tensor<i13>
// CHECK-ON: }
//
// CHECK-OFF-LABEL: func.func @sparse_reduction_ori_accumulator_on_rhs(
// CHECK-OFF-SAME: %[[VAL_0:.*]]: tensor<i13>,
// CHECK-OFF-SAME: %[[VAL_1:.*]]: tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>>) -> tensor<i13> {
// CHECK-OFF-DAG: %[[VAL_2:.*]] = arith.constant 0 : index
// CHECK-OFF-DAG: %[[VAL_3:.*]] = arith.constant 1 : index
// CHECK-OFF: %[[VAL_4:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
// CHECK-OFF: %[[VAL_5:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?xi13, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xi13>
// CHECK-OFF: %[[VAL_6:.*]] = bufferization.to_memref %[[VAL_0]] : memref<i13>
// CHECK-OFF: %[[VAL_7:.*]] = memref.load %[[VAL_6]][] : memref<i13>
// CHECK-OFF: %[[VAL_8:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_2]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_9:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_3]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_10:.*]] = scf.for %[[VAL_11:.*]] = %[[VAL_8]] to %[[VAL_9]] step %[[VAL_3]] iter_args(%[[VAL_12:.*]] = %[[VAL_7]]) -> (i13) {
// CHECK-OFF: %[[VAL_13:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_11]]] : memref<?xi13>
// CHECK-OFF: %[[VAL_14:.*]] = arith.ori %[[VAL_13]], %[[VAL_12]] : i13
// CHECK-OFF: scf.yield %[[VAL_14]] : i13
// CHECK-OFF: } {"Emitted from" = "linalg.generic"}
// CHECK-OFF: memref.store %[[VAL_15:.*]], %[[VAL_6]][] : memref<i13>
// CHECK-OFF: %[[VAL_16:.*]] = bufferization.to_tensor %[[VAL_6]] : memref<i13>
// CHECK-OFF: return %[[VAL_16]] : tensor<i13>
// CHECK-OFF: }
#SparseVector = #sparse_tensor.encoding<{dimLevelType = ["compressed"]}>
#trait = {
indexing_maps = [
affine_map<(i) -> (i)>, // a (in)
affine_map<(i) -> ()> // x (out)
],
iterator_types = ["reduction"]
}
func.func @sparse_reduction_ori_accumulator_on_rhs(%argx: tensor<i13>,
%arga: tensor<?xi13, #SparseVector>)
-> tensor<i13> {
%0 = linalg.generic #trait
ins(%arga: tensor<?xi13, #SparseVector>)
outs(%argx: tensor<i13>) {
^bb(%a: i13, %x: i13):
%t = arith.ori %a, %x: i13
linalg.yield %t : i13
} -> tensor<i13>
return %0 : tensor<i13>
}
// -----
// Check that we vectorize reduction with subi.
//
aartbikUnsubmitted
Done
ReplyInline Actions

wer -> we

aartbik: wer -> we
qcolombetAuthorUnsubmitted
Done
ReplyInline Actions

Good catch

qcolombet: Good catch
// CHECK-ON-LABEL: func.func @sparse_reduction_subi(
// CHECK-ON-SAME: %[[VAL_0:.*]]: tensor<i32>,
// CHECK-ON-SAME: %[[VAL_1:.*]]: tensor<?xi32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>>) -> tensor<i32> {
// CHECK-ON-DAG: %[[VAL_2:.*]] = arith.constant 8 : index
// CHECK-ON-DAG: %[[VAL_3:.*]] = arith.constant 0 : index
// CHECK-ON-DAG: %[[VAL_4:.*]] = arith.constant dense<0> : vector<8xi32>
// CHECK-ON-DAG: %[[VAL_5:.*]] = arith.constant 1 : index
// CHECK-ON: %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<?xi32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
// CHECK-ON: %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?xi32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xi32>
// CHECK-ON: %[[VAL_8:.*]] = bufferization.to_memref %[[VAL_0]] : memref<i32>
// CHECK-ON: %[[VAL_9:.*]] = memref.load %[[VAL_8]][] : memref<i32>
// CHECK-ON: %[[VAL_10:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_3]]] : memref<?xindex>
// CHECK-ON: %[[VAL_11:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_5]]] : memref<?xindex>
// CHECK-ON: %[[VAL_12:.*]] = vector.insertelement %[[VAL_9]], %[[VAL_4]]{{\[}}%[[VAL_3]] : index] : vector<8xi32>
// CHECK-ON: %[[VAL_13:.*]] = scf.for %[[VAL_14:.*]] = %[[VAL_10]] to %[[VAL_11]] step %[[VAL_2]] iter_args(%[[VAL_15:.*]] = %[[VAL_12]]) -> (vector<8xi32>) {
// CHECK-ON: %[[VAL_16:.*]] = affine.min #map(%[[VAL_11]], %[[VAL_14]]){{\[}}%[[VAL_2]]]
// CHECK-ON: %[[VAL_17:.*]] = vector.create_mask %[[VAL_16]] : vector<8xi1>
// CHECK-ON: %[[VAL_18:.*]] = vector.maskedload %[[VAL_7]]{{\[}}%[[VAL_14]]], %[[VAL_17]], %[[VAL_4]] : memref<?xi32>, vector<8xi1>, vector<8xi32> into vector<8xi32>
// CHECK-ON: %[[VAL_19:.*]] = arith.subi %[[VAL_15]], %[[VAL_18]] : vector<8xi32>
// CHECK-ON: %[[VAL_20:.*]] = arith.select %[[VAL_17]], %[[VAL_19]], %[[VAL_15]] : vector<8xi1>, vector<8xi32>
// CHECK-ON: scf.yield %[[VAL_20]] : vector<8xi32>
// CHECK-ON: } {"Emitted from" = "linalg.generic"}
// CHECK-ON: %[[VAL_21:.*]] = vector.reduction <add>, %[[VAL_22:.*]] : vector<8xi32> into i32
// CHECK-ON: memref.store %[[VAL_21]], %[[VAL_8]][] : memref<i32>
// CHECK-ON: %[[VAL_23:.*]] = bufferization.to_tensor %[[VAL_8]] : memref<i32>
// CHECK-ON: return %[[VAL_23]] : tensor<i32>
// CHECK-ON: }
//
// CHECK-OFF-LABEL: func.func @sparse_reduction_subi(
// CHECK-OFF-SAME: %[[VAL_0:.*]]: tensor<i32>,
// CHECK-OFF-SAME: %[[VAL_1:.*]]: tensor<?xi32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>>) -> tensor<i32> {
// CHECK-OFF-DAG: %[[VAL_2:.*]] = arith.constant 0 : index
// CHECK-OFF-DAG: %[[VAL_3:.*]] = arith.constant 1 : index
// CHECK-OFF: %[[VAL_4:.*]] = sparse_tensor.pointers %[[VAL_1]] {dimension = 0 : index} : tensor<?xi32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xindex>
// CHECK-OFF: %[[VAL_5:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<?xi32, #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>> to memref<?xi32>
// CHECK-OFF: %[[VAL_6:.*]] = bufferization.to_memref %[[VAL_0]] : memref<i32>
// CHECK-OFF: %[[VAL_7:.*]] = memref.load %[[VAL_6]][] : memref<i32>
// CHECK-OFF: %[[VAL_8:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_2]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_9:.*]] = memref.load %[[VAL_4]]{{\[}}%[[VAL_3]]] : memref<?xindex>
// CHECK-OFF: %[[VAL_10:.*]] = scf.for %[[VAL_11:.*]] = %[[VAL_8]] to %[[VAL_9]] step %[[VAL_3]] iter_args(%[[VAL_12:.*]] = %[[VAL_7]]) -> (i32) {
// CHECK-OFF: %[[VAL_13:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_11]]] : memref<?xi32>
// CHECK-OFF: %[[VAL_14:.*]] = arith.subi %[[VAL_12]], %[[VAL_13]] : i32
// CHECK-OFF: scf.yield %[[VAL_14]] : i32
// CHECK-OFF: } {"Emitted from" = "linalg.generic"}
// CHECK-OFF: memref.store %[[VAL_15:.*]], %[[VAL_6]][] : memref<i32>
// CHECK-OFF: %[[VAL_16:.*]] = bufferization.to_tensor %[[VAL_6]] : memref<i32>
// CHECK-OFF: return %[[VAL_16]] : tensor<i32>
// CHECK-OFF: }
#SparseVector = #sparse_tensor.encoding<{dimLevelType = ["compressed"]}>
#trait = {
indexing_maps = [
affine_map<(i) -> (i)>, // a (in)
affine_map<(i) -> ()> // x (out)
],
iterator_types = ["reduction"]
}
func.func @sparse_reduction_subi(%argx: tensor<i32>,
%arga: tensor<?xi32, #SparseVector>)
-> tensor<i32> {
%0 = linalg.generic #trait
ins(%arga: tensor<?xi32, #SparseVector>)
outs(%argx: tensor<i32>) {
^bb(%a: i32, %x: i32):
%t = arith.subi %x, %a: i32
linalg.yield %t : i32
} -> tensor<i32>
return %0 : tensor<i32>
}
// -----
aartbikUnsubmitted
Done
ReplyInline Actions

can you remove this part from the test now?

aartbik: can you remove this part from the test now?
qcolombetAuthorUnsubmitted
Done
ReplyInline Actions

Sure thing!
I’ll do that next week

qcolombet: Sure thing! I’ll do that next week