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.

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

[mlir][vector] Constrain patterns: vector.contract -> vector.outerproduct #68400

Closed
Closed
Show file tree
Hide file tree
Changes from 1 commit
Commits
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
57 changes: 39 additions & 18 deletions mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -424,9 +424,14 @@ struct UnrolledOuterProductGenerator
return rewriter.create<arith::ExtSIOp>(loc, promotedType, v);
}

FailureOr<Value> outerProd(Value lhs, Value rhs, Value res, int reductionSize,
FailureOr<Value> outerProd(Value lhs, Value rhs, Value res, VectorType lhsType,
int reductionDim,
std::optional<Value> maybeMask = std::nullopt) {
assert(reductionSize > 0);
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

why was this assert removed?

// Unrolling a scalable dimension would be incorrect - bail out.
if (lhsType.getScalableDims()[reductionDim])
return failure();

int reductionSize = lhsType.getDimSize(reductionDim);
// Incremental support for masking.
if (mask && !maybeMask.has_value())
return failure();
Expand Down Expand Up @@ -459,33 +464,39 @@ struct UnrolledOuterProductGenerator
Value transposedMask = t(mask, {2, 0, 1});
// Classical row-major matmul: Just permute the lhs.
if (layout({{m, k}, {k, n}, {m, n}}))
return outerProd(t(lhs), rhs, res, lhsType.getDimSize(1), transposedMask);
return outerProd(t(lhs), rhs, res, lhsType, /*reductionDim=*/1,
transposedMask);
// TODO: may be better to fail and use some vector<k> -> scalar reduction.
if (layout({{m, k}, {n, k}, {m, n}})) {
Value tlhs = t(lhs);
return outerProd(tlhs, t(rhs), res, lhsType.getDimSize(1),
return outerProd(tlhs, t(rhs), res, lhsType, /*reductionDim=*/1,
transposedMask);
}
// No need to permute anything.
if (layout({{k, m}, {k, n}, {m, n}}))
return outerProd(lhs, rhs, res, lhsType.getDimSize(0), transposedMask);
return outerProd(lhs, rhs, res, lhsType, /*reductionDim=*/0,
transposedMask);
// Just permute the rhs.
if (layout({{k, m}, {n, k}, {m, n}}))
return outerProd(lhs, t(rhs), res, lhsType.getDimSize(0), transposedMask);
return outerProd(lhs, t(rhs), res, lhsType, /*reductionDim=*/0,
transposedMask);
// Transposed output: swap RHS and LHS.
// Classical row-major matmul: permute the lhs.
if (layout({{m, k}, {k, n}, {n, m}}))
return outerProd(rhs, t(lhs), res, lhsType.getDimSize(1), transposedMask);
return outerProd(rhs, t(lhs), res, lhsType, /*reductionDim=*/1,
transposedMask);
// TODO: may be better to fail and use some vector<k> -> scalar reduction.
if (layout({{m, k}, {n, k}, {n, m}})) {
Value trhs = t(rhs);
return outerProd(trhs, t(lhs), res, lhsType.getDimSize(1),
return outerProd(trhs, t(lhs), res, lhsType, /*reductionDim=*/1,
transposedMask);
}
if (layout({{k, m}, {k, n}, {n, m}}))
return outerProd(rhs, lhs, res, lhsType.getDimSize(0), transposedMask);
return outerProd(rhs, lhs, res, lhsType, /*reductionDim=*/0,
transposedMask);
if (layout({{k, m}, {n, k}, {n, m}}))
return outerProd(t(rhs), lhs, res, lhsType.getDimSize(0), transposedMask);
return outerProd(t(rhs), lhs, res, lhsType, /*reductionDim=*/0,
transposedMask);
return failure();
}

Expand All @@ -503,16 +514,20 @@ struct UnrolledOuterProductGenerator

// Case mat-vec: transpose.
if (layout({{m, k}, {k}, {m}}))
return outerProd(t(lhs), rhs, res, lhsType.getDimSize(1), transposedMask);
return outerProd(t(lhs), rhs, res, lhsType, /*reductionDim=*/1,
transposedMask);
// Case mat-trans-vec: ready to go.
if (layout({{k, m}, {k}, {m}}))
return outerProd(lhs, rhs, res, lhsType.getDimSize(0), transposedMask);
return outerProd(lhs, rhs, res, lhsType, /*reductionDim=*/0,
transposedMask);
// Case vec-mat: swap and transpose.
if (layout({{k}, {m, k}, {m}}))
return outerProd(t(rhs), lhs, res, lhsType.getDimSize(0), transposedMask);
return outerProd(t(rhs), lhs, res, lhsType, /*reductionDim=*/0,
transposedMask);
// Case vec-mat-trans: swap and ready to go.
if (layout({{k}, {k, m}, {m}}))
return outerProd(rhs, lhs, res, lhsType.getDimSize(0), transposedMask);
return outerProd(rhs, lhs, res, lhsType, /*reductionDim=*/0,
transposedMask);
return failure();
}

Expand All @@ -528,16 +543,16 @@ struct UnrolledOuterProductGenerator

// Case mat-vec: transpose.
if (layout({{m, k}, {k}, {m}}))
return outerProd(t(lhs), rhs, res, lhsType.getDimSize(1), mask);
return outerProd(t(lhs), rhs, res, lhsType, /*reductionDim=*/1, mask);
// Case mat-trans-vec: ready to go.
if (layout({{k, m}, {k}, {m}}))
return outerProd(lhs, rhs, res, lhsType.getDimSize(0), mask);
return outerProd(lhs, rhs, res, lhsType, /*reductionDim=*/0, mask);
// Case vec-mat: swap and transpose.
if (layout({{k}, {m, k}, {m}}))
return outerProd(t(rhs), lhs, res, lhsType.getDimSize(0), mask);
return outerProd(t(rhs), lhs, res, lhsType, /*reductionDim=*/0, mask);
// Case vec-mat-trans: swap and ready to go.
if (layout({{k}, {k, m}, {m}}))
return outerProd(rhs, lhs, res, lhsType.getDimSize(0), mask);
return outerProd(rhs, lhs, res, lhsType, /*reductionDim=*/0, mask);
return failure();
}

Expand Down Expand Up @@ -980,9 +995,15 @@ FailureOr<Value> ContractionOpLowering::lowerParallel(PatternRewriter &rewriter,
diag << "expected lhsIndex=" << lhsIndex << " and rhsIndex=" << rhsIndex
<< " to map to the same dimension";
});
// Unrolling a scalable dimension would be incorrect - bail out.
if (lhsType.getScalableDims()[lhsIndex])
return failure();
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can we notify with a message here and below ?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, will do :)

dimSize = lhsType.getDimSize(lhsIndex);
} else if (rhsIndex >= 0) {
iterIndex = iMap[1].getDimPosition(rhsIndex);
// Unrolling a scalable dimension would be incorrect - bail out.
if (rhsType.getScalableDims()[rhsIndex])
return failure();
dimSize = rhsType.getDimSize(rhsIndex);
}
if (iterIndex < 0)
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,33 @@
// RUN: mlir-opt %s --test-transform-dialect-interpreter -allow-unregistered-dialect --split-input-file --verify-diagnostics

#matvec_accesses = [
affine_map<(i, j) -> (i, j)>,
affine_map<(i, j) -> (j)>,
affine_map<(i, j) -> (i)>
]
#matvec_trait = {
indexing_maps = #matvec_accesses,
iterator_types = ["parallel", "reduction"]
}

// Unrolling scalable reduction dim is not supported - bail out

// expected-error@below {{greedy pattern application failed}}
func.func @masked_extract_contract2_scalable_reduction_dim(%arg0: vector<[2]x[3]xf32>,
%arg1: vector<[3]xf32>,
%arg2: vector<[2]xf32>,
%m: vector<[2]x[3]xi1>) -> vector<[2]xf32> {
%0 = vector.mask %m { vector.contract #matvec_trait %arg0, %arg1, %arg2
: vector<[2]x[3]xf32>, vector<[3]xf32> into vector<[2]xf32> } : vector<[2]x[3]xi1> -> vector<[2]xf32>
return %0 : vector<[2]xf32>
}

transform.sequence failures(propagate) {
^bb1(%module_op: !transform.any_op):
%f = transform.structured.match ops{["func.func"]} in %module_op
: (!transform.any_op) -> !transform.any_op

transform.apply_patterns to %f {
transform.apply_patterns.vector.lower_contraction lowering_strategy = "outerproduct"
} : !transform.any_op
}
Original file line number Diff line number Diff line change
Expand Up @@ -31,19 +31,19 @@
}

// CHECK-LABEL: func.func @masked_extract_contract2(
// CHECK-SAME: %[[VAL_0:.*]]: vector<2x3xf32>,
// CHECK-SAME: %[[VAL_1:.*]]: vector<3xf32>,
// CHECK-SAME: %[[VAL_2:.*]]: vector<2xf32>,
// CHECK-SAME: %[[IN_MASK:.*]]: vector<2x3xi1>) -> vector<2xf32>
// CHECK-SAME: %{{.*}}: vector<2x3xf32>,
// CHECK-SAME: %{{.*}}: vector<3xf32>,
// CHECK-SAME: %{{.*}}: vector<2xf32>,
// CHECK-SAME: %[[IN_MASK:.*]]: vector<2x3xi1>) -> vector<2xf32>
// CHECK: %[[T_MASK:.*]] = vector.transpose %[[IN_MASK]], [1, 0] : vector<2x3xi1> to vector<3x2xi1>
// CHECK: %[[MASK0:.*]] = vector.extract %[[T_MASK]][0] : vector<2xi1> from vector<3x2xi1>
// CHECK: vector.mask %[[MASK0]] { vector.outerproduct
// CHECK: vector.mask %[[MASK0]] { vector.outerproduct {{.*}} {kind = #vector.kind<add>} : vector<2xf32>, f32 } : vector<2xi1> -> vector<2xf32>

// CHECK: %[[MASK1:.*]] = vector.extract %[[T_MASK]][1] : vector<2xi1> from vector<3x2xi1>
// CHECK: vector.mask %[[MASK1]] { vector.outerproduct
// CHECK: vector.mask %[[MASK1]] { vector.outerproduct {{.*}} {kind = #vector.kind<add>} : vector<2xf32>, f32 } : vector<2xi1> -> vector<2xf32>

// CHECK: %[[MASK2:.*]] = vector.extract %[[T_MASK]][2] : vector<2xi1> from vector<3x2xi1>
// CHECK: vector.mask %[[MASK2]] { vector.outerproduct
// CHECK: vector.mask %[[MASK2]] { vector.outerproduct {{.*}} {kind = #vector.kind<add>} : vector<2xf32>, f32 } : vector<2xi1> -> vector<2xf32>

func.func @masked_extract_contract2(%arg0: vector<2x3xf32>,
%arg1: vector<3xf32>,
Expand All @@ -54,22 +54,29 @@ func.func @masked_extract_contract2(%arg0: vector<2x3xf32>,
return %0 : vector<2xf32>
}

// CHECK-LABEL: func.func @masked_extract_contract4(
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why was this fixed size test removed?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copy and paste failure, sorry :( Thanks for pointing this out!

// CHECK-SAME: %[[VAL_0:.*]]: vector<3x5xf32>,
// CHECK-SAME: %[[VAL_1:.*]]: vector<5x7xf32>,
// CHECK-SAME: %[[VAL_2:.*]]: vector<3x7xf32>,
// CHECK-SAME: %[[VAL_3:.*]]: vector<3x7x5xi1>) -> vector<3x7xf32> {
// CHECK: %[[VAL_5:.*]] = vector.transpose %[[VAL_3]], [2, 0, 1] : vector<3x7x5xi1> to vector<5x3x7xi1>
// CHECK: %[[VAL_8:.*]] = vector.extract %[[VAL_5]][0] : vector<3x7xi1> from vector<5x3x7xi1>
// CHECK: %[[VAL_9:.*]] = vector.mask %[[VAL_8]] { vector.outerproduct %{{.*}} {kind = #vector.kind<add>} : vector<3xf32>, vector<7xf32> } : vector<3x7xi1> -> vector<3x7xf32>
// CHECK: %[[VAL_12:.*]] = vector.extract %[[VAL_5]][1] : vector<3x7xi1> from vector<5x3x7xi1>
// CHECK: %[[VAL_13:.*]] = vector.mask %[[VAL_12]] { vector.outerproduct %{{.*}} {kind = #vector.kind<add>} : vector<3xf32>, vector<7xf32> } : vector<3x7xi1> -> vector<3x7xf32>
// CHECK: %[[VAL_16:.*]] = vector.extract %[[VAL_5]][2] : vector<3x7xi1> from vector<5x3x7xi1>
// CHECK: %[[VAL_17:.*]] = vector.mask %[[VAL_16]] { vector.outerproduct %{{.*}} {kind = #vector.kind<add>} : vector<3xf32>, vector<7xf32> } : vector<3x7xi1> -> vector<3x7xf32>
// CHECK: %[[VAL_20:.*]] = vector.extract %[[VAL_5]][3] : vector<3x7xi1> from vector<5x3x7xi1>
// CHECK: %[[VAL_21:.*]] = vector.mask %[[VAL_20]] { vector.outerproduct %{{.*}} {kind = #vector.kind<add>} : vector<3xf32>, vector<7xf32> } : vector<3x7xi1> -> vector<3x7xf32>
// CHECK: %[[VAL_24:.*]] = vector.extract %[[VAL_5]][4] : vector<3x7xi1> from vector<5x3x7xi1>
// CHECK: %[[VAL_25:.*]] = vector.mask %[[VAL_24]] { vector.outerproduct %{{.*}} {kind = #vector.kind<add>} : vector<3xf32>, vector<7xf32> } : vector<3x7xi1> -> vector<3x7xf32>

// CHECK-LABEL: func.func @masked_extract_contract2_scalable_parallel_dim(
// CHECK-SAME: %{{.*}}: vector<[2]x3xf32>,
// CHECK-SAME: %{{.*}}: vector<3xf32>,
// CHECK-SAME: %{{.*}}: vector<[2]xf32>,
// CHECK-SAME: %[[IN_MASK:.*]]: vector<[2]x3xi1>) -> vector<[2]xf32>
// CHECK: %[[T_MASK:.*]] = vector.transpose %[[IN_MASK]], [1, 0] : vector<[2]x3xi1> to vector<3x[2]xi1>
// CHECK: %[[MASK0:.*]] = vector.extract %[[T_MASK]][0] : vector<[2]xi1> from vector<3x[2]xi1>
// CHECK: vector.mask %[[MASK0]] { vector.outerproduct {{.*}} {kind = #vector.kind<add>} : vector<[2]xf32>, f32 } : vector<[2]xi1> -> vector<[2]xf32>

// CHECK: %[[MASK1:.*]] = vector.extract %[[T_MASK]][1] : vector<[2]xi1> from vector<3x[2]xi1>
// CHECK: vector.mask %[[MASK1]] { vector.outerproduct {{.*}} {kind = #vector.kind<add>} : vector<[2]xf32>, f32 } : vector<[2]xi1> -> vector<[2]xf32>

// CHECK: %[[MASK2:.*]] = vector.extract %[[T_MASK]][2] : vector<[2]xi1> from vector<3x[2]xi1>
// CHECK: vector.mask %[[MASK2]] { vector.outerproduct {{.*}} {kind = #vector.kind<add>} : vector<[2]xf32>, f32 } : vector<[2]xi1> -> vector<[2]xf32>
func.func @masked_extract_contract2_scalable_parallel_dim(%arg0: vector<[2]x3xf32>,
%arg1: vector<3xf32>,
%arg2: vector<[2]xf32>,
%m: vector<[2]x3xi1>) -> vector<[2]xf32> {
%0 = vector.mask %m { vector.contract #matvec_trait %arg0, %arg1, %arg2
: vector<[2]x3xf32>, vector<3xf32> into vector<[2]xf32> } : vector<[2]x3xi1> -> vector<[2]xf32>
return %0 : vector<[2]xf32>
}

func.func @masked_extract_contract4(%arg0: vector<3x5xf32>,
%arg1: vector<5x7xf32>,
Expand Down