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

Reland "Support fusing broadcast transposes with attention" #19962

Merged
merged 2 commits into from
Feb 12, 2025
+88 −24
Merged

Reland "Support fusing broadcast transposes with attention" #19962

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
10c8b1b
Revert "Revert "Support fusing broadcast transposes with attention" (…
IanWood1 Feb 11, 2025
d624520
Fix attn info
IanWood1 Feb 11, 2025
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
43 changes: 23 additions & 20 deletions compiler/src/iree/compiler/Dialect/LinalgExt/Transforms/TransposeFusion.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@
#include "iree/compiler/Dialect/LinalgExt/Transforms/Transforms.h"
#include "iree/compiler/Dialect/LinalgExt/Utils/IndexingUtils.h"
#include "llvm/ADT/STLExtras.h"
#include "mlir/Dialect/Linalg/IR/LinalgInterfaces.h"
#include "mlir/Dialect/Linalg/Transforms/Transforms.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/Dialect/Utils/StructuredOpsUtils.h"
Expand Down Expand Up @@ -58,42 +59,44 @@ struct FuseTransposeWithAttentionOp final

LogicalResult matchAndRewrite(LinalgExt::AttentionOp attentionOp,
PatternRewriter &rewriter) const override {
OpOperand *transposeOperand = nullptr;
linalg::LinalgOp transposeOp;
OpOperand *operand = nullptr;
linalg::LinalgOp producer;
for (OpOperand *input : attentionOp.getDpsInputOperands()) {
if (controlFn && !controlFn(input)) {
continue;
}

auto maybeTransposeOp = input->get().getDefiningOp<linalg::LinalgOp>();
if (maybeTransposeOp && isaTranspose(maybeTransposeOp) &&
maybeTransposeOp->hasOneUse()) {
transposeOp = maybeTransposeOp;
transposeOperand = input;
auto maybeProducer = input->get().getDefiningOp<linalg::GenericOp>();
if (maybeProducer && maybeProducer.isSingleYieldOp()) {
producer = maybeProducer;
operand = input;
break;
}
}
if (!transposeOperand) {
return rewriter.notifyMatchFailure(attentionOp, "no transpose operand");
if (!operand) {
return rewriter.notifyMatchFailure(attentionOp, "no operand found");
}

int64_t inputIndex = transposeOperand->getOperandNumber();
SmallVector<int64_t> perm = getPermutation(transposeOp);
auto invPerm = invertPermutationVector(perm);
int64_t inputIndex = operand->getOperandNumber();

auto producerMaps = producer.getIndexingMapsArray();
AffineMap producerInputMap = producerMaps[0];
AffineMap producerResultMap = producerMaps[1];
if (!producerInputMap.isProjectedPermutation() ||
!producerResultMap.isPermutation()) {
return failure();
}

rewriter.modifyOpInPlace(attentionOp, [&]() {
SmallVector<AffineMap> newIndexingMaps =
attentionOp.getIndexingMapsArray();
AffineMap inputMap = attentionOp.getMatchingIndexingMap(transposeOperand);
SmallVector<AffineExpr> newExprs =
applyPermutation(inputMap.getResults(), invPerm);
AffineMap transposedMap =
AffineMap::get(inputMap.getNumDims(), inputMap.getNumSymbols(),
newExprs, rewriter.getContext());
newIndexingMaps[inputIndex] = transposedMap;
AffineMap consumerInputMap = attentionOp.getMatchingIndexingMap(operand);
AffineMap composedMap =
producerInputMap.compose(inversePermutation(producerResultMap));
newIndexingMaps[inputIndex] = composedMap.compose(consumerInputMap);
attentionOp.setIndexingMapsAttr(
rewriter.getAffineMapArrayAttr(newIndexingMaps));
attentionOp.setOperand(inputIndex, transposeOp.getDpsInputs()[0]);
attentionOp.setOperand(inputIndex, producer.getDpsInputs()[0]);
});

return success();
Expand Down
6 changes: 2 additions & 4 deletions compiler/src/iree/compiler/Dialect/LinalgExt/Utils/IndexingUtils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,12 +37,10 @@ void AttentionOpDetail::inferFromIndexingMaps(AffineMap qMap, AffineMap kMap,
llvm::SmallDenseSet<int64_t> vSet = findPermutationsIndexingOperand(vMap);
llvm::SmallDenseSet<int64_t> oSet = findPermutationsIndexingOperand(oMap);

// B = (Q & K & O) U (K & V & O)
// B = (Q & V) U (K & O)
Copy link
Contributor

Choose a reason for hiding this comment

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

Interesting. Is this true. cc @Groverkss

Copy link
Contributor

Choose a reason for hiding this comment

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

Yes. We can actually make it just (Q & V) or just (K & O) also because they are the same.

K := batch x k1 x k2
O := batch x m x n

so their intersection has to be the batch dimensions.

Similarily,

Q := batch x m x k1
V := batch x k2 x n

so their intersection has to be the batch dimension

Interestingly, this actually made me realize we are doing GQA wrong. GQA is a broadcast on the M dimension, which is why it's only on Q.

llvm::SmallDenseSet<int64_t> b1Set = qSet;
llvm::set_intersect(b1Set, kSet);
llvm::set_intersect(b1Set, oSet);
llvm::set_intersect(b1Set, vSet);
llvm::SmallDenseSet<int64_t> b2Set = kSet;
llvm::set_intersect(b2Set, vSet);
llvm::set_intersect(b2Set, oSet);
llvm::SmallDenseSet<int64_t> bSet = b1Set;
llvm::set_union(bSet, b2Set);
Expand Down
63 changes: 63 additions & 0 deletions compiler/src/iree/compiler/DispatchCreation/test/elementwise_op_fusion.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -208,6 +208,8 @@ util.func public @fuse_generic_gather2(
// CHECK-NEXT: %[[RES4:[a-zA-Z0-9]+]] = arith.addf %[[RES2]], %[[RES3]] : f32
// CHECK-NEXT: linalg.yield %[[RES4]] : f32

// -----

util.func public @fuse_transpose_attention_to_producer(%q: tensor<2x10x4096x64xf16>, %k: tensor<2x10x4096x64xf16>, %quantized_v: tensor<2x10x4096x64xi32>, %quant_offset: tensor<10x64xi32>, %quant_scale: tensor<10x64xf32>, %scale: f16) -> tensor<2x10x4096x64xf16> {
// Dequantize int-quantization of V
%init_dequant = tensor.empty() : tensor<2x10x4096x64xf16>
Expand Down Expand Up @@ -258,3 +260,64 @@ util.func public @fuse_transpose_attention_to_producer(%q: tensor<2x10x4096x64xf
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5) -> ()>
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5) -> (d0, d1, d2, d5)>
// CHECK-SAME: ins(%[[ARG0]], %[[ARG1]], %[[DEQUANT_V]], %[[ARG5]]

// -----

util.func public @fuse_attention_with_broadcast(%arg0: tensor<4x8x128x?xf16>, %arg1: tensor<4x8x4x?x32x128xf16>, %arg2: tensor<4x8x4x?x128xf16>, %arg3: f16, %arg4: tensor<4x8x4x?x32x?xf16>, %arg5: tensor<4x8x4x?x32x128xf16>, %arg6: tensor<4x8x4x128x?xf16>) -> tensor<4x8x4x?x32x128xf16> {
%0 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d3, d4)>, affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%arg0 : tensor<4x8x128x?xf16>) outs(%arg6 : tensor<4x8x4x128x?xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<4x8x4x128x?xf16>
%1 = iree_linalg_ext.attention {indexing_maps = [affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d6)>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d7, d6)>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d5, d7)>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> ()>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d7)>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5)>]} ins(%arg1, %arg2, %0, %arg3, %arg4 : tensor<4x8x4x?x32x128xf16>, tensor<4x8x4x?x128xf16>, tensor<4x8x4x128x?xf16>, f16, tensor<4x8x4x?x32x?xf16>) outs(%arg5 : tensor<4x8x4x?x32x128xf16>) {
^bb0(%arg7: f32):
iree_linalg_ext.yield %arg7 : f32
} -> tensor<4x8x4x?x32x128xf16>
util.return %1 : tensor<4x8x4x?x32x128xf16>
}

// CHECK-LABEL: func public @fuse_attention_with_broadcast
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]:
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]:
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]:
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9]+]]:
// CHECK-SAME: %[[ARG4:[a-zA-Z0-9]+]]:
// CHECK: %[[ATTENTION:.+]] = iree_linalg_ext.attention
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d6)>,
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d7, d6)>,
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d5, d7)>,
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> ()>,
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d7)>
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5)>
// CHECK-SAME: ins(%[[ARG1]], %[[ARG2]], %[[ARG0]], %[[ARG3]], %[[ARG4]] :
// CHECK: util.return %[[ATTENTION]]


// -----

util.func public @fuse_attention_with_broadcast_transpose(%arg0: tensor<4x?x8x128xf16>, %arg1: tensor<4x8x4x?x32x128xf16>, %arg2: tensor<4x8x4x?x128xf16>, %arg3: f16, %arg4: tensor<4x8x4x?x32x?xf16>, %arg5: tensor<4x8x4x?x32x128xf16>, %arg6: tensor<4x8x4x128x?xf16>) -> tensor<4x8x4x?x32x128xf16> {
%0 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d4)>, affine_map<(d0, d1, d2, d3, d4) -> (d0, d2, d3, d4, d1)>], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%arg0 : tensor<4x?x8x128xf16>) outs(%arg6 : tensor<4x8x4x128x?xf16>) {
^bb0(%in: f16, %out: f16):
linalg.yield %in : f16
} -> tensor<4x8x4x128x?xf16>
%1 = iree_linalg_ext.attention {indexing_maps = [affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d6)>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d7, d6)>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d5, d7)>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> ()>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d7)>, affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5)>]} ins(%arg1, %arg2, %0, %arg3, %arg4 : tensor<4x8x4x?x32x128xf16>, tensor<4x8x4x?x128xf16>, tensor<4x8x4x128x?xf16>, f16, tensor<4x8x4x?x32x?xf16>) outs(%arg5 : tensor<4x8x4x?x32x128xf16>) {
^bb0(%arg7: f32):
iree_linalg_ext.yield %arg7 : f32
} -> tensor<4x8x4x?x32x128xf16>
util.return %1 : tensor<4x8x4x?x32x128xf16>
}

// CHECK-LABEL: func public @fuse_attention_with_broadcast_transpose
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]:
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]:
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]:
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9]+]]:
// CHECK-SAME: %[[ARG4:[a-zA-Z0-9]+]]:
// CHECK: %[[ATTENTION:.+]] = iree_linalg_ext.attention
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d6)>,
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d7, d6)>,
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d7, d1, d5)>,
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> ()>,
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d7)>
// CHECK-SAME: affine_map<(d0, d1, d2, d3, d4, d5, d6, d7) -> (d0, d1, d2, d3, d4, d5)>
// CHECK-SAME: ins(%[[ARG1]], %[[ARG2]], %[[ARG0]], %[[ARG3]], %[[ARG4]] :
// CHECK: util.return %[[ATTENTION]]
Loading