Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Dispatch] Disable UnpackLikeOp+ExtractSlice fusion. #19408

Merged
merged 1 commit into from
Dec 8, 2024
+34 −55
Merged

[Dispatch] Disable UnpackLikeOp+ExtractSlice fusion. #19408

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 4 additions & 34 deletions compiler/src/iree/compiler/DispatchCreation/FormDispatchRegions.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -182,28 +182,9 @@ static bool isPackLikeOp(Operation *op) {
return isa<IREE::Encoding::SetEncodingOp, tensor::PackOp>(op);
}

/// Returns true if the operation is an `unpack` op or an `unset_encoding` op,
/// or an `extract_slice` op whose source operand matches those criteria,
/// recursively.
/// The idea is that we want to ensure that `extract_slice` ops can't prevent
/// fusion between a `unset_encoding` producer and some linalg consumer. In
/// %0 = unset_encoding ...
/// %1 = extract_slice %0 ...
/// %2 = linalg.generic ins(%1) ...
/// we are not content to be fusing %1 into %0, we also want to be fusing %2,
/// so we want to prevent %1 from acting as a consumer fusion barrier.
static bool isUnpackLikeOpViaExtractSliceOps(Operation *op) {
if (isa<IREE::Encoding::UnsetEncodingOp, tensor::UnPackOp>(op)) {
return true;
}
if (isa<tensor::ExtractSliceOp>(op)) {
Value source = op->getOperand(0);
Operation *producer = source.getDefiningOp();
if (isUnpackLikeOpViaExtractSliceOps(producer)) {
return true;
}
}
return false;
/// Returns true if the operation is an `unpack` op or an `unset_encoding` op.
static bool isUnpackLikeOp(Operation *op) {
return isa<IREE::Encoding::UnsetEncodingOp, tensor::UnPackOp>(op);
}

/// Since `iree_encoding.set_encoding` doesnt have padding semantics a
Expand Down Expand Up @@ -476,18 +457,7 @@ isFusableWithConsumer(OpOperand &fusedOperand,

// Fuse unset_encoding operations with `tensor.extract_slice` and elementwise
// generic ops.
if (isUnpackLikeOpViaExtractSliceOps(producer)) {
// Fuse `unset_encoding` -> `extract_slice` op since they get folded into
// `unpack` on materialization.
if (isa<tensor::ExtractSliceOp>(consumer)) {
auto sliceOp = cast<tensor::ExtractSliceOp>(consumer);
return llvm::all_of(
sliceOp.getMixedOffsets(),
[](OpFoldResult ofr) { return isConstantIntValue(ofr, 0); }) &&
llvm::all_of(sliceOp.getMixedStrides(), [](OpFoldResult ofr) {
return isConstantIntValue(ofr, 1);
});
}
if (isUnpackLikeOp(producer)) {
// Fuse `unset_encoding/unpack` -> elementwise operations. Fuse unpack with
// non-overlapping reductions (i.e., the reduction dimension is not packed).
if (auto consumerLinalgOp = dyn_cast<linalg::LinalgOp>(consumer)) {
Expand Down
11 changes: 4 additions & 7 deletions compiler/src/iree/compiler/DispatchCreation/test/dispatch_linalg_on_tensors.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1958,17 +1958,15 @@ util.func public @pad_and_set_encoding_op(%arg0 : tensor<?x?xf32>)
// -----

#encoding = #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>
util.func public @unset_encoding_and_slice(
util.func public @unset_encoding_with_encoded_slice(
%arg0: tensor<?x?xf32, #encoding>,
%arg1 : index, %arg2 : index) -> tensor<?x?xf32> {
%0 = iree_encoding.unset_encoding %arg0
: tensor<?x?xf32, #encoding> -> tensor<?x?xf32>{%arg1, %arg2}
%1 = tensor.extract_slice %0[0, 0] [%arg1, %arg2] [1, 1]
: tensor<?x?xf32> to tensor<?x?xf32>
util.return %1 : tensor<?x?xf32>
util.return %0 : tensor<?x?xf32>
}
// CHECK: #[[ENCODING:.+]] = #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>
// CHECK: util.func public @unset_encoding_and_slice
// CHECK: util.func public @unset_encoding_with_encoded_slice
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x?xf32, #[[ENCODING]]>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]: index
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]: index
Expand All @@ -1991,8 +1989,7 @@ util.func public @unset_encoding_and_slice(
// CHECK-SAME: sizes = [%[[D0_W]], %[[D1_W]]]
// CHECK-SAME: !flow.dispatch.tensor<readonly:tensor<?x?xf32, #[[ENCODING]]>>{%[[D0_W]], %[[D1_W]]}
// CHECK: %[[UNSET_ENCODING:.+]] = iree_encoding.unset_encoding %[[LOAD]]
// CHECK: %[[SLICE:.+]] = tensor.extract_slice %[[UNSET_ENCODING]][0, 0] [%[[ARG0_W]], %[[ARG1_W]]]
// CHECK: flow.dispatch.tensor.store %[[SLICE]], %[[OUTARG]]
// CHECK: flow.dispatch.tensor.store %[[UNSET_ENCODING]], %[[OUTARG]]
// CHECK-SAME: sizes = [%[[ARG0_W]], %[[ARG1_W]]]
// CHECK-SAME: !flow.dispatch.tensor<writeonly:tensor<?x?xf32>>{%[[ARG0_W]], %[[ARG1_W]]}
// CHECK: flow.return
Expand Down
40 changes: 26 additions & 14 deletions compiler/src/iree/compiler/DispatchCreation/test/form_dispatch_regions.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -304,38 +304,34 @@ util.func public @unset_encoding_elementwise_fusion(
// -----

#encoding = #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>
util.func public @unset_encoding_slice_elementwise_fusion(
util.func public @unset_encoding_elementwise_fusion(
%arg0: tensor<?x?xf32, #encoding>,
%arg1: tensor<?xf32>, %arg2 : index, %arg3 : index) -> tensor<?x?xf32> {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%0 = iree_encoding.unset_encoding %arg0
: tensor<?x?xf32, #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>> -> tensor<?x?xf32>{%arg2, %arg3}
%1 = tensor.extract_slice %0[0, 0] [%arg2, %arg3] [1, 1] : tensor<?x?xf32> to tensor<?x?xf32>
%2 = tensor.dim %1, %c0 : tensor<?x?xf32>
%3 = tensor.dim %1, %c1 : tensor<?x?xf32>
%4 = tensor.empty(%2, %3) : tensor<?x?xf32>
%5 = linalg.generic {
%1 = tensor.empty(%arg2, %arg3) : tensor<?x?xf32>
%2 = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
affine_map<(d0, d1) -> (d0)>,
affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]}
ins(%1, %arg1 : tensor<?x?xf32>, tensor<?xf32>)
outs(%4 : tensor<?x?xf32>) {
ins(%0, %arg1 : tensor<?x?xf32>, tensor<?xf32>)
outs(%1 : tensor<?x?xf32>) {
^bb0(%b0 : f32, %b1 : f32, %b2 : f32):
%6 = arith.addf %b0, %b1 : f32
linalg.yield %6 : f32
%3 = arith.addf %b0, %b1 : f32
linalg.yield %3 : f32
} -> tensor<?x?xf32>
util.return %5 : tensor<?x?xf32>
util.return %2 : tensor<?x?xf32>
}
// CHECK: #[[$ENCODING:.+]] = #iree_encoding.encoding<operand_index = 0 : i64, op_type = matmul, element_types = [f32, f32, f32]>
// CHECK-LABEL: util.func public @unset_encoding_slice_elementwise_fusion(
// CHECK-LABEL: util.func public @unset_encoding_elementwise_fusion(
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x?xf32, #[[$ENCODING]]>
// CHECK-SAME: %[[ARG1:.+]]: tensor<?xf32>
// CHECK: %[[RESULT0:.+]] = flow.dispatch.region
// CHECK: %[[UNSET_ENCODING:.+]] = iree_encoding.unset_encoding %[[ARG0]]
// CHECK: %[[SLICE:.+]] = tensor.extract_slice %[[UNSET_ENCODING]]
// CHECK: %[[GENERIC:.+]] = linalg.generic {{.*}} ins(%[[SLICE]]
// CHECK: %[[GENERIC:.+]] = linalg.generic {{.*}} ins(%[[UNSET_ENCODING]]
// CHECK: flow.return %[[GENERIC]]
// CHECK: util.return %[[RESULT0]]

Expand Down Expand Up @@ -382,6 +378,22 @@ util.func public @unpack_elementwise_fusion(

// -----

#encoding = #iree_encoding.encoding<operand_index = 2 : index, op_type = matmul, element_types = [f32, f32, f32]>
util.func public @unset_encoding_slice(%arg0: tensor<1x50x384xf32, #encoding>) -> tensor<384xf32> {
%0 = iree_encoding.unset_encoding %arg0 : tensor<1x50x384xf32, #encoding> -> tensor<1x50x384xf32>
%extracted_slice = tensor.extract_slice %0[0, 0, 0] [1, 1, 384] [1, 1, 1] : tensor<1x50x384xf32> to tensor<384xf32>
util.return %extracted_slice : tensor<384xf32>
}
// CHECK-LABEL: util.func public @unset_encoding_slice
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]
// CHECK: %[[RESULT:.+]] = flow.dispatch.region
// CHECK: %[[UNSET_ENCODING:.+]] = iree_encoding.unset_encoding
// CHECK: flow.return %[[UNSET_ENCODING]]
// CHECK: %[[SLICE:.+]] = tensor.extract_slice %[[RESULT]]
// CHECK: util.return %[[SLICE]]

// -----

util.func public @unpack_non_intersecting_reduction(
%arg0: tensor<?x?x?xf32>,
%arg1: tensor<?xf32>) -> tensor<?xf32> {
Expand Down
Loading