Skip to content

[mlir][sparse] fix error when convolution stride is applied on a dens… #79521

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

Merged
merged 1 commit into from
Jan 26, 2024
Merged

[mlir][sparse] fix error when convolution stride is applied on a dens… #79521

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
4 changes: 2 additions & 2 deletions mlir/lib/Dialect/SparseTensor/Transforms/Utils/LoopEmitter.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -313,8 +313,8 @@ void LoopEmitter::initSubSectIterator(OpBuilder &builder, Location loc) {
Value loopHi = loopHighs[loop];
size = ADDI(size, MULI(loopHi, C_IDX(stride)));
}
it = makeNonEmptySubSectIterator(builder, loc, parent, std::move(lvlIt),
size, curDep.second);
it = makeNonEmptySubSectIterator(builder, loc, parent, loopHighs[loop],
std::move(lvlIt), size, curDep.second);
} else {
Value size = loopHighs[loop];
const SparseIterator &subSectIter = *iters[t][lvl].back();
Expand Down
9 changes: 6 additions & 3 deletions mlir/lib/Dialect/SparseTensor/Transforms/Utils/SparseTensorLevel.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1271,17 +1271,20 @@ static const IterType *unwrapFilter(const SparseIterator *it) {
}

std::unique_ptr<SparseIterator> sparse_tensor::makeNonEmptySubSectIterator(
OpBuilder &b, Location l, const SparseIterator *parent,
OpBuilder &b, Location l, const SparseIterator *parent, Value loopBound,
std::unique_ptr<SparseIterator> &&delegate, Value size, unsigned stride) {

// Try unwrap the NonEmptySubSectIterator from a filter parent.
parent = tryUnwrapFilter<NonEmptySubSectIterator>(parent);
auto it = std::make_unique<NonEmptySubSectIterator>(
b, l, parent, std::move(delegate), size);

if (stride != 1)
if (stride != 1) {
// TODO: We can safely skip bound checking on sparse levels, but for dense
// iteration space, we need the bound to infer the dense loop range.
return std::make_unique<FilterIterator>(std::move(it), /*offset=*/C_IDX(0),
C_IDX(stride), /*size=*/C_IDX(-1));
C_IDX(stride), /*size=*/loopBound);
}
return it;
}

Expand Down
2 changes: 1 addition & 1 deletion mlir/lib/Dialect/SparseTensor/Transforms/Utils/SparseTensorLevel.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -246,7 +246,7 @@ makeSlicedLevelIterator(std::unique_ptr<SparseIterator> &&sit, Value offset,
/// Helper function to create a SparseIterator object that iterate over the
/// non-empty subsections set.
std::unique_ptr<SparseIterator> makeNonEmptySubSectIterator(
OpBuilder &b, Location l, const SparseIterator *parent,
OpBuilder &b, Location l, const SparseIterator *parent, Value loopBound,
std::unique_ptr<SparseIterator> &&delegate, Value size, unsigned stride);

/// Helper function to create a SparseIterator object that iterate over a
Expand Down
21 changes: 21 additions & 0 deletions mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_strided_conv_2d_nhwc_hwcf.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,6 +69,14 @@ func.func @conv_2d_nhwc_hwcf_CDCC(%arg0: tensor<?x?x?x?xf32, #CDCC>, %arg1: tens
return %ret : tensor<?x?x?x?xf32>
}

func.func @conv_2d_nhwc_hwcf_dual_CDCC(%arg0: tensor<?x?x?x?xf32, #CDCC>, %arg1: tensor<?x?x?x?xf32, #CDCC>, %arg2: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32> {
%ret = linalg.conv_2d_nhwc_hwcf {dilations = dense<1> : tensor<2xi64>,
strides = dense<2> : tensor<2xi64>}
ins (%arg0, %arg1: tensor<?x?x?x?xf32, #CDCC>, tensor<?x?x?x?xf32, #CDCC>)
outs (%arg2: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
return %ret : tensor<?x?x?x?xf32>
}


func.func @entry() {
%c0 = arith.constant 0 : index
Expand All @@ -87,16 +95,28 @@ func.func @entry() {

%in2D_nhwc_CCCC = sparse_tensor.convert %in2D_nhwc
: tensor<?x?x?x?xf32> to tensor<?x?x?x?xf32, #CCCC>
%filter2D_nhwc_CDCC = sparse_tensor.convert %filter2D_nhwc
: tensor<?x?x?x?xf32> to tensor<?x?x?x?xf32, #CDCC>
%in2D_nhwc_CDCC = sparse_tensor.convert %in2D_nhwc
: tensor<?x?x?x?xf32> to tensor<?x?x?x?xf32, #CDCC>

%dense_ret = call @conv_2d_nhwc_hwcf(%in2D_nhwc, %filter2D_nhwc, %out2D_nhwc) : (tensor<?x?x?x?xf32>, tensor<?x?x?x?xf32>, tensor<?x?x?x?xf32>) -> (tensor<?x?x?x?xf32>)
%CCCC_ret = call @conv_2d_nhwc_hwcf_CCCC(%in2D_nhwc_CCCC, %filter2D_nhwc, %out2D_nhwc) : (tensor<?x?x?x?xf32, #CCCC>, tensor<?x?x?x?xf32>, tensor<?x?x?x?xf32>) -> (tensor<?x?x?x?xf32>)
%CDCC_ret = call @conv_2d_nhwc_hwcf_CDCC(%in2D_nhwc_CDCC, %filter2D_nhwc, %out2D_nhwc) : (tensor<?x?x?x?xf32, #CDCC>, tensor<?x?x?x?xf32>, tensor<?x?x?x?xf32>) -> (tensor<?x?x?x?xf32>)

%dual_CDCC_ret = call @conv_2d_nhwc_hwcf_dual_CDCC(%in2D_nhwc_CDCC, %filter2D_nhwc_CDCC, %out2D_nhwc)
: (tensor<?x?x?x?xf32, #CDCC>, tensor<?x?x?x?xf32, #CDCC>, tensor<?x?x?x?xf32>) -> (tensor<?x?x?x?xf32>)

// CHECK: ( ( ( ( 0 ), ( 0 ), ( 0 ) ), ( ( 0 ), ( 20 ), ( 0 ) ), ( ( 0 ), ( 0 ), ( 0 ) ) ),
// CHECK-SAME: ( ( ( 0 ), ( 0 ), ( 0 ) ), ( ( 0 ), ( 0 ), ( 0 ) ), ( ( 0 ), ( 0 ), ( 0 ) ) ),
// CHECK-SAME: ( ( ( 0 ), ( 0 ), ( 0 ) ), ( ( 0 ), ( 0 ), ( 0 ) ), ( ( 0 ), ( 0 ), ( 0 ) ) ) )
%v_dual = vector.transfer_read %dual_CDCC_ret[%c0, %c0, %c0, %c0], %zero
: tensor<?x?x?x?xf32>, vector<3x3x3x1xf32>
vector.print %v_dual : vector<3x3x3x1xf32>

// CHECK-NEXT: ( ( ( ( 0 ), ( 0 ), ( 0 ) ), ( ( 0 ), ( 20 ), ( 0 ) ), ( ( 0 ), ( 0 ), ( 0 ) ) ),
// CHECK-SAME: ( ( ( 0 ), ( 0 ), ( 0 ) ), ( ( 0 ), ( 0 ), ( 0 ) ), ( ( 0 ), ( 0 ), ( 0 ) ) ),
// CHECK-SAME: ( ( ( 0 ), ( 0 ), ( 0 ) ), ( ( 0 ), ( 0 ), ( 0 ) ), ( ( 0 ), ( 0 ), ( 0 ) ) ) )
%dense_v = vector.transfer_read %dense_ret[%c0, %c0, %c0, %c0], %zero
: tensor<?x?x?x?xf32>, vector<3x3x3x1xf32>
vector.print %dense_v : vector<3x3x3x1xf32>
Expand All @@ -120,6 +140,7 @@ func.func @entry() {
bufferization.dealloc_tensor %filter2D_nhwc : tensor<?x?x?x?xf32>
bufferization.dealloc_tensor %out2D_nhwc : tensor<?x?x?x?xf32>

bufferization.dealloc_tensor %filter2D_nhwc_CDCC : tensor<?x?x?x?xf32, #CDCC>
bufferization.dealloc_tensor %in2D_nhwc_CCCC : tensor<?x?x?x?xf32, #CCCC>
bufferization.dealloc_tensor %in2D_nhwc_CDCC : tensor<?x?x?x?xf32, #CDCC>
return
Expand Down