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

[NFC][Im2Col] Move im2tocall to affine.delinearize_index #19840

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[NFC][Im2Col] Move im2tocall to affine.delinearize_index #19840

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
22 changes: 8 additions & 14 deletions compiler/src/iree/compiler/Dialect/LinalgExt/IR/AggregatedOpInterfaceImpl.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -786,13 +786,10 @@ FailureOr<SmallVector<Value>> Im2colOp::decomposeOperation(OpBuilder &b) {
OpFoldResult ivOffset = mulOfrs(b, nestedLoc, stride, ivs[ivIdx]);
kIndex = addOfrs(b, nestedLoc, kIndex, ivOffset);
}
FailureOr<SmallVector<Value>> maybeDelinKOffset = affine::delinearizeIndex(
b, nestedLoc, getValueOrCreateConstantIndexOp(b, loc, kIndex),
getValueOrCreateConstantIndexOp(b, loc, (kBasis)));
if (failed(maybeDelinKOffset)) {
return failure();
}
SmallVector<Value> delinKOffset = maybeDelinKOffset.value();
auto delinKOffsetOp = b.create<affine::AffineDelinearizeIndexOp>(
nestedLoc, getValueOrCreateConstantIndexOp(b, loc, kIndex), kBasis,
/*hasOuterBound=*/true);
SmallVector<Value> delinKOffset = delinKOffsetOp.getResults();
// Split the delinearized offsets into the window offsets (for M offsets)
// and the K offsets for the input tensor.
SmallVector<Value> windowOffset, inputKOffset;
Expand Down Expand Up @@ -823,13 +820,10 @@ FailureOr<SmallVector<Value>> Im2colOp::decomposeOperation(OpBuilder &b) {
// Delinearize the m_offset * m_strides into the convolution output space.
// `mBasis` contains the basis for the iteration space of result of the
// convolution op (i.e., basis for result H and W dims).
FailureOr<SmallVector<Value>> maybeDelinMOffset = affine::delinearizeIndex(
b, nestedLoc,
getValueOrCreateConstantIndexOp(b, nestedLoc, linearMOffset), mBasis);
if (failed(maybeDelinMOffset)) {
return failure();
}
SmallVector<Value> delinMOffset = maybeDelinMOffset.value();
auto delinMOffsetOp = b.create<affine::AffineDelinearizeIndexOp>(
nestedLoc, getValueOrCreateConstantIndexOp(b, loc, linearMOffset), mBasis,
/*hasOuterBound=*/true);
SmallVector<Value> delinMOffset = delinMOffsetOp.getResults();

// Compute the final offsets into the input tensor.
OpFoldResult zero = b.getIndexAttr(0);
Expand Down
90 changes: 53 additions & 37 deletions compiler/src/iree/compiler/Dialect/LinalgExt/Transforms/test/decompose_im2col.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,9 +15,8 @@ module {
return %7 : tensor<2x?x4xf32>
}
}
// CHECK-DAG: #[[$MAP:.+]] = affine_map<()[s0] -> (s0 * 4 - (s0 floordiv 160) * 640)>
// CHECK-DAG: #[[$MAP1:.+]] = affine_map<(d0)[s0, s1] -> ((d0 + s0) floordiv 32 + s1 floordiv 480)>
// CHECK-DAG: #[[$MAP2:.+]] = affine_map<(d0)[s0, s1] -> ((d0 + s0) mod 32 + s1 floordiv 160 - (s1 floordiv 480) * 3)>
// CHECK-DAG: #[[$MAP:.+]] = affine_map<()[s0] -> (s0 * 4)>
// CHECK-DAG: #[[$MAP1:.+]] = affine_map<(d0)[s0] -> (d0 + s0)>
// CHECK-LABEL: func.func @im2col_untile_k
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]
// CHECK-SAME: %[[mSIZE:[a-zA-Z0-9_]+]]
Expand All @@ -27,12 +26,15 @@ module {
// CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index
// CHECK-DAG: %[[C2:.+]] = arith.constant 2 : index
// CHECK: %[[OUT_TILE:.+]] = tensor.empty(%[[mSIZE]]) : tensor<2x?x4xf32>
// CHECK: %[[kScaled:.+]] = affine.apply #[[$MAP]]()[%[[K]]]
// CHECK: %[[bLOOP:.+]] = scf.for %[[b:.+]] = %[[C0]] to %[[C2]] step %[[C1]] iter_args(%[[OUT0:.+]] = %[[OUT_TILE]]) -> (tensor<2x?x4xf32>)
// CHECK: %[[mLOOP:.+]] = scf.for %[[m:.+]] = %[[C0]] to %[[mSIZE]] step %[[C1]] iter_args(%[[OUT1:.+]] = %[[OUT0]]) -> (tensor<2x?x4xf32>)
// CHECK-DAG: %[[kIDX:.+]] = affine.apply #[[$MAP]]()[%[[K]]]
// CHECK-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[m]])[%[[mOFF]], %[[K]]]
// CHECK-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP2]](%[[m]])[%[[mOFF]], %[[K]]]
// CHECK: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[b]], %[[hIDX]], %[[wIDX]], %[[kIDX]]] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x4xf32>
// CHECK-DAG: %[[kParts:.+]]:3 = affine.delinearize_index %[[kScaled]] into (3, 3, 640)
// CHECK-DAG: %[[mIDX:.+]] = affine.apply #[[$MAP1]](%[[m]])[%[[mOFF]]]
// CHECK-DAG: %[[mParts:.+]]:2 = affine.delinearize_index %[[mIDX]] into (32, 32)
// CHECK-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[mParts]]#0)[%[[kParts]]#0]
// CHECK-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP1]](%[[mParts]]#1)[%[[kParts]]#1]
// CHECK: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[b]], %[[hIDX]], %[[wIDX]], %[[kParts]]#2] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x4xf32>
// CHECK: %[[OUT_SLICE:.+]] = tensor.extract_slice %[[OUT1]][%[[b]], %[[m]], 0] [1, 1, 4] [1, 1, 1] : tensor<2x?x4xf32> to tensor<1x1x4xf32>
// CHECK: %[[COPY:.+]] = linalg.copy ins(%[[IN_SLICE]] : tensor<1x1x4xf32>) outs(%[[OUT_SLICE]] : tensor<1x1x4xf32>) -> tensor<1x1x4xf32>
// CHECK: %[[INSERT:.+]] = tensor.insert_slice %[[COPY]] into %[[OUT1]][%[[b]], %[[m]], 0] [1, 1, 4] [1, 1, 1] : tensor<1x1x4xf32> into tensor<2x?x4xf32>
Expand All @@ -57,9 +59,9 @@ module {
return %8 : tensor<2x?x?xf32>
}
}
// CHECK-DAG: #[[$MAP:.+]] = affine_map<(d0)[s0] -> ((d0 + s0) floordiv 10)>
// CHECK-DAG: #[[$MAP1:.+]] = affine_map<(d0, d1)[s0, s1] -> (((d0 + s0) floordiv 32) * 5 + (((d1 + s1) mod 10) floordiv 5) * 4)>
// CHECK-DAG: #[[$MAP2:.+]] = affine_map<(d0, d1)[s0, s1] -> (d0 * 3 + d1 * 7 + s0 * 3 + s1 * 7 - ((d0 + s0) floordiv 32) * 96 - ((d1 + s1) floordiv 5) * 35)>
// CHECK-DAG: #[[$MAP:.+]] = affine_map<(d0)[s0] -> (d0 + s0)>
// CHECK-DAG: #[[$MAP1:.+]] = affine_map<(d0, d1) -> (d0 * 5 + d1 * 4)>
// CHECK-DAG: #[[$MAP2:.+]] = affine_map<(d0, d1) -> (d0 * 3 + d1 * 7)>
// CHECK-LABEL: func.func @im2col_transposed_m_pos
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]
// CHECK-SAME: %[[mSIZE:[a-zA-Z0-9_]+]]
Expand All @@ -74,9 +76,12 @@ module {
// CHECK: %[[mLOOP:.+]] = scf.for %[[m:.+]] = %[[C0]] to %[[mSIZE]] step %[[C1]] iter_args(%[[OUT1:.+]] = %[[OUT0]]) -> (tensor<2x?x?xf32>)
// CHECK: %[[kLOOP:.+]] = scf.for %[[k:.+]] = %[[C0]] to %[[kSIZE]] step %[[C1]] iter_args(%[[OUT2:.+]] = %[[OUT1]]) -> (tensor<2x?x?xf32>)
// CHECK-DAG: %[[kIDX:.+]] = affine.apply #[[$MAP]](%[[k]])[%[[kOFF]]]
// CHECK-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[m]], %[[k]])[%[[mOFF]], %[[kOFF]]]
// CHECK-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP2]](%[[m]], %[[k]])[%[[mOFF]], %[[kOFF]]]
// CHECK: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[kIDX]], %[[b]], %[[wIDX]], %[[hIDX]]] [1, 1, 1, 1] [1, 1, 1, 1] : tensor<640x2x101x172xf32> to tensor<1x1x1xf32>
// CHECK-DAG: %[[kParts:.+]]:3 = affine.delinearize_index %[[kIDX]] into (640, 2, 5)
// CHECK-DAG: %[[mIDX:.+]] = affine.apply #[[$MAP]](%[[m]])[%[[mOFF]]]
// CHECK-DAG: %[[mParts:.+]]:2 = affine.delinearize_index %[[mIDX]] into (32, 32)
// CHECK-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[mParts]]#0, %[[kParts]]#1)
// CHECK-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP2]](%[[mParts]]#1, %[[kParts]]#2)
// CHECK: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[kParts]]#0, %[[b]], %[[wIDX]], %[[hIDX]]] [1, 1, 1, 1] [1, 1, 1, 1] : tensor<640x2x101x172xf32> to tensor<1x1x1xf32>
// CHECK: %[[OUT_SLICE:.+]] = tensor.extract_slice %[[OUT2]][%[[b]], %[[m]], %[[k]]] [1, 1, 1] [1, 1, 1] : tensor<2x?x?xf32> to tensor<1x1x1xf32>
// CHECK: %[[COPY:.+]] = linalg.copy ins(%[[IN_SLICE]] : tensor<1x1x1xf32>) outs(%[[OUT_SLICE]] : tensor<1x1x1xf32>) -> tensor<1x1x1xf32>
// CHECK: %[[INSERT:.+]] = tensor.insert_slice %[[COPY]] into %[[OUT2]][%[[b]], %[[m]], %[[k]]] [1, 1, 1] [1, 1, 1] : tensor<1x1x1xf32> into tensor<2x?x?xf32>
Expand All @@ -99,9 +104,9 @@ module {
return %7 : tensor<2x?x?x2x4xf32>
}
}
// CHECK-DAG: #[[$MAP:.+]] = affine_map<(d0)[s0] -> (d0 * 4 + s0 * 4 - ((d0 + s0) floordiv 160) * 640)>
// CHECK-DAG: #[[$MAP1:.+]] = affine_map<(d0, d1, d2)[s0, s1, s2, s3] -> ((d1 + s2 + d0 * s0 + s1 * s0) floordiv 32 + (d2 + s3) floordiv 480)>
// CHECK-DAG: #[[$MAP2:.+]] = affine_map<(d0, d1, d2)[s0, s1, s2, s3] -> (d0 * s0 + d1 + s1 * s0 + s2 - ((d1 + s2 + d0 * s0 + s1 * s0) floordiv 32) * 32 + (d2 + s3) floordiv 160 - ((d2 + s3) floordiv 480) * 3)>
// CHECK-DAG: #[[$MAP:.+]] = affine_map<(d0)[s0] -> (d0 * 4 + s0 * 4)
// CHECK-DAG: #[[$MAP1:.+]] = affine_map<(d0, d1)[s0, s1, s2] -> (d0 * s0 + d1 + s1 * s0 + s2)>
// CHECK-DAG: #[[$MAP2:.+]] = affine_map<(d0, d1) -> (d0 + d1)>
// CHECK-LABEL: func.func @im2col_expanded
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]
// CHECK-SAME: %[[mSIZE0:[a-zA-Z0-9_]+]]
Expand All @@ -119,9 +124,12 @@ module {
// CHECK: %[[mLOOP1:.+]] = scf.for %[[m1:.+]] = %[[C0]] to %[[mSIZE1]] step %[[C1]] iter_args(%[[OUT2:.+]] = %[[OUT1]]) -> (tensor<2x?x?x2x4xf32>)
// CHECK: %[[kLOOP:.+]] = scf.for %[[k:.+]] = %[[C0]] to %[[C2]] step %[[C1]] iter_args(%[[OUT3:.+]] = %[[OUT2]]) -> (tensor<2x?x?x2x4xf32>)
// CHECK-DAG: %[[kIDX:.+]] = affine.apply #[[$MAP]](%[[k]])[%[[kOFF]]]
// CHECK-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[m0]], %[[m1]], %[[k]])[%[[mSTRIDE]], %[[mOFF0]], %[[mOFF1]], %[[kOFF]]]
// CHECK-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP2]](%[[m0]], %[[m1]], %[[k]])[%[[mSTRIDE]], %[[mOFF0]], %[[mOFF1]], %[[kOFF]]]
// CHECK: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[b]], %[[hIDX]], %[[wIDX]], %[[kIDX]]] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x1x4xf32>
// CHECK-DAG: %[[kParts:.+]]:3 = affine.delinearize_index %[[kIDX]] into (3, 3, 640)
// CHECK-DAG: %[[mIDX:.+]] = affine.apply #[[$MAP1]](%[[m0]], %[[m1]])[%[[mSTRIDE]], %[[mOFF0]], %[[mOFF1]]]
// CHECK-DAG: %[[mParts:.+]]:2 = affine.delinearize_index %[[mIDX]] into (32, 32)
// CHECK-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP2]](%[[mParts]]#0, %[[kParts]]#0)
// CHECK-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP2]](%[[mParts]]#1, %[[kParts]]#1)
// CHECK: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[b]], %[[hIDX]], %[[wIDX]], %[[kParts]]#2] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x1x4xf32>
// CHECK: %[[OUT_SLICE:.+]] = tensor.extract_slice %[[OUT3]][%[[b]], %[[m0]], %[[m1]], %[[k]], 0] [1, 1, 1, 1, 4] [1, 1, 1, 1, 1] : tensor<2x?x?x2x4xf32> to tensor<1x1x1x4xf32>
// CHECK: %[[COPY:.+]] = linalg.copy ins(%[[IN_SLICE]] : tensor<1x1x1x4xf32>) outs(%[[OUT_SLICE]] : tensor<1x1x1x4xf32>) -> tensor<1x1x1x4xf32>
// CHECK: %[[INSERT:.+]] = tensor.insert_slice %[[COPY]] into %[[OUT3]][%[[b]], %[[m0]], %[[m1]], %[[k]], 0] [1, 1, 1, 1, 4] [1, 1, 1, 1, 1] : tensor<1x1x1x4xf32> into tensor<2x?x?x2x4xf32>
Expand Down Expand Up @@ -165,9 +173,8 @@ module {
return %7 : tensor<2x2x4xf32>
}
}
// CHECK-UNROLL-DAG: #[[$MAP:.+]] = affine_map<()[s0] -> (s0 * 4 - (s0 floordiv 160) * 640)>
// CHECK-UNROLL-DAG: #[[$MAP1:.+]] = affine_map<(d0)[s0, s1] -> ((d0 + s0) floordiv 32 + s1 floordiv 480)>
// CHECK-UNROLL-DAG: #[[$MAP2:.+]] = affine_map<(d0)[s0, s1] -> ((d0 + s0) mod 32 + s1 floordiv 160 - (s1 floordiv 480) * 3)>
// CHECK-UNROLL-DAG: #[[$MAP:.+]] = affine_map<()[s0] -> (s0 * 4)>
// CHECK-UNROLL-DAG: #[[$MAP1:.+]] = affine_map<(d0)[s0] -> (d0 + s0)>
// CHECK-UNROLL-LABEL: func.func @im2col_unrolled
// CHECK-UNROLL-SAME: %[[ARG0:[a-zA-Z0-9_]+]]
// CHECK-UNROLL-SAME: %[[mOFF:[a-zA-Z0-9_]+]]
Expand All @@ -179,39 +186,48 @@ module {
// First iteration
//
// CHECK-UNROLL-DAG: %[[kIDX:.+]] = affine.apply #[[$MAP]]()[%[[K]]]
// CHECK-UNROLL-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[C0]])[%[[mOFF]], %[[K]]]
// CHECK-UNROLL-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP2]](%[[C0]])[%[[mOFF]], %[[K]]]
// CHECK-UNROLL: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[C0]], %[[hIDX]], %[[wIDX]], %[[kIDX]]] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL-DAG: %[[kParts:.+]]:3 = affine.delinearize_index %[[kIDX]] into (3, 3, 640)
// CHECK-UNROLL-DAG: %[[mIDX:.+]] = affine.apply #[[$MAP1]](%[[C0]])[%[[mOFF]]]
// CHECK-UNROLL-DAG: %[[mParts:.+]]:2 = affine.delinearize_index %[[mIDX]] into (32, 32)
// CHECK-UNROLL-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[mParts]]#0)[%[[kParts]]#0]
// CHECK-UNROLL-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP1]](%[[mParts]]#1)[%[[kParts]]#1]
// CHECK-UNROLL: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[C0]], %[[hIDX]], %[[wIDX]], %[[kParts]]#2] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL: %[[OUT_SLICE:.+]] = tensor.extract_slice %[[OUT_TILE]][%[[C0]], %[[C0]], 0] [1, 1, 4] [1, 1, 1] : tensor<2x2x4xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL: %[[COPY:.+]] = linalg.copy ins(%[[IN_SLICE]] : tensor<1x1x4xf32>) outs(%[[OUT_SLICE]] : tensor<1x1x4xf32>) -> tensor<1x1x4xf32>
// CHECK-UNROLL: %[[INSERT0:.+]] = tensor.insert_slice %[[COPY]] into %[[OUT_TILE]][%[[C0]], %[[C0]], 0] [1, 1, 4] [1, 1, 1] : tensor<1x1x4xf32> into tensor<2x2x4xf32>

// Second iteration
//
// CHECK-UNROLL-DAG: %[[kIDX:.+]] = affine.apply #[[$MAP]]()[%[[K]]]
// CHECK-UNROLL-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[C1]])[%[[mOFF]], %[[K]]]
// CHECK-UNROLL-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP2]](%[[C1]])[%[[mOFF]], %[[K]]]
// CHECK-UNROLL: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[C0]], %[[hIDX]], %[[wIDX]], %[[kIDX]]] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL-DAG: %[[kParts:.+]]:3 = affine.delinearize_index %[[kIDX]] into (3, 3, 640)
// CHECK-UNROLL-DAG: %[[mIDX:.+]] = affine.apply #[[$MAP1]](%[[C1]])[%[[mOFF]]]
// CHECK-UNROLL-DAG: %[[mParts:.+]]:2 = affine.delinearize_index %[[mIDX]] into (32, 32)
// CHECK-UNROLL-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[mParts]]#0)[%[[kParts]]#0]
// CHECK-UNROLL-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP1]](%[[mParts]]#1)[%[[kParts]]#1]
// CHECK-UNROLL: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[C0]], %[[hIDX]], %[[wIDX]], %[[kParts]]#2] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL: %[[OUT_SLICE:.+]] = tensor.extract_slice %[[INSERT0]][%[[C0]], %[[C1]], 0] [1, 1, 4] [1, 1, 1] : tensor<2x2x4xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL: %[[COPY:.+]] = linalg.copy ins(%[[IN_SLICE]] : tensor<1x1x4xf32>) outs(%[[OUT_SLICE]] : tensor<1x1x4xf32>) -> tensor<1x1x4xf32>
// CHECK-UNROLL: %[[INSERT1:.+]] = tensor.insert_slice %[[COPY]] into %[[INSERT0]][%[[C0]], %[[C1]], 0] [1, 1, 4] [1, 1, 1] : tensor<1x1x4xf32> into tensor<2x2x4xf32>

// Third iteration
//
// CHECK-UNROLL-DAG: %[[kIDX:.+]] = affine.apply #[[$MAP]]()[%[[K]]]
// CHECK-UNROLL-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[C0]])[%[[mOFF]], %[[K]]]
// CHECK-UNROLL-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP2]](%[[C0]])[%[[mOFF]], %[[K]]]
// CHECK-UNROLL: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[C1]], %[[hIDX]], %[[wIDX]], %[[kIDX]]] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL-DAG: %[[kParts:.+]]:3 = affine.delinearize_index %[[kIDX]] into (3, 3, 640)
// CHECK-UNROLL-DAG: %[[mIDX:.+]] = affine.apply #[[$MAP1]](%[[C0]])[%[[mOFF]]]
// CHECK-UNROLL-DAG: %[[mParts:.+]]:2 = affine.delinearize_index %[[mIDX]] into (32, 32)
// CHECK-UNROLL-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[mParts]]#0)[%[[kParts]]#0]
// CHECK-UNROLL-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP1]](%[[mParts]]#1)[%[[kParts]]#1]
// CHECK-UNROLL: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[C1]], %[[hIDX]], %[[wIDX]], %[[kParts]]#2] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL: %[[OUT_SLICE:.+]] = tensor.extract_slice %[[INSERT1]][%[[C1]], %[[C0]], 0] [1, 1, 4] [1, 1, 1] : tensor<2x2x4xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL: %[[COPY:.+]] = linalg.copy ins(%[[IN_SLICE]] : tensor<1x1x4xf32>) outs(%[[OUT_SLICE]] : tensor<1x1x4xf32>) -> tensor<1x1x4xf32>
// CHECK-UNROLL: %[[INSERT2:.+]] = tensor.insert_slice %[[COPY]] into %[[INSERT1]][%[[C1]], %[[C0]], 0] [1, 1, 4] [1, 1, 1] : tensor<1x1x4xf32> into tensor<2x2x4xf32>

// Fourth iteration
//
// CHECK-UNROLL-DAG: %[[kIDX:.+]] = affine.apply #[[$MAP]]()[%[[K]]]
// CHECK-UNROLL-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[C1]])[%[[mOFF]], %[[K]]]
// CHECK-UNROLL-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP2]](%[[C1]])[%[[mOFF]], %[[K]]]
// CHECK-UNROLL: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[C1]], %[[hIDX]], %[[wIDX]], %[[kIDX]]] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL-DAG: %[[kParts:.+]]:3 = affine.delinearize_index %[[kIDX]] into (3, 3, 640)
// CHECK-UNROLL-DAG: %[[mIDX:.+]] = affine.apply #[[$MAP1]](%[[C1]])[%[[mOFF]]]
// CHECK-UNROLL-DAG: %[[mParts:.+]]:2 = affine.delinearize_index %[[mIDX]] into (32, 32)
// CHECK-UNROLL-DAG: %[[hIDX:.+]] = affine.apply #[[$MAP1]](%[[mParts]]#0)[%[[kParts]]#0]
// CHECK-UNROLL-DAG: %[[wIDX:.+]] = affine.apply #[[$MAP1]](%[[mParts]]#1)[%[[kParts]]#1]
// CHECK-UNROLL: %[[IN_SLICE:.+]] = tensor.extract_slice %[[ARG0]][%[[C1]], %[[hIDX]], %[[wIDX]], %[[kParts]]#2] [1, 1, 1, 4] [1, 1, 1, 1] : tensor<2x34x34x640xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL: %[[OUT_SLICE:.+]] = tensor.extract_slice %[[INSERT2]][%[[C1]], %[[C1]], 0] [1, 1, 4] [1, 1, 1] : tensor<2x2x4xf32> to tensor<1x1x4xf32>
// CHECK-UNROLL: %[[COPY:.+]] = linalg.copy ins(%[[IN_SLICE]] : tensor<1x1x4xf32>) outs(%[[OUT_SLICE]] : tensor<1x1x4xf32>) -> tensor<1x1x4xf32>
// CHECK-UNROLL: %[[INSERT3:.+]] = tensor.insert_slice %[[COPY]] into %[[INSERT2]][%[[C1]], %[[C1]], 0] [1, 1, 4] [1, 1, 1] : tensor<1x1x4xf32> into tensor<2x2x4xf32>
Expand Down
Loading