[Transform] Add basic onednn_graph dialect lowering #61
Conversation
LongshengDu
force-pushed
the
longsheng/add_onednn_graph
branch
from
4e629e5 to
cd43162
Compare
LongshengDu
force-pushed
the
longsheng/onednn_lower
branch
from
5ee1278 to
c06f84d
Compare
| int64_t rank = resultTy.getRank(); | ||
| SmallVector<int64_t> permutation(rank); | ||
| std::iota(std::begin(permutation), std::end(permutation), 0); | ||
| permutation[rank - 2] = rank - 1; | ||
| permutation[rank - 1] = rank - 2; |
There was a problem hiding this comment.
Since this already assumes it's 2d, can it be static?
There was a problem hiding this comment.
Yeah, for 2D matmul it should be static
| } | ||
| } | ||
|
|
||
| typedef Value (*GetOperandFn)(Operation *, PatternRewriter &, TensorType); |
There was a problem hiding this comment.
The definition is a bit confusing since there's mlir::Value (even though the result is still the same). If there's any better option - go for it. Otherwise it's fine.
There was a problem hiding this comment.
This is a type define for a function that returns mlir::Value.
| struct OriginalOperand { | ||
| template <unsigned I> | ||
| static Value getIdx(Operation *op, PatternRewriter &b, TensorType ty) { | ||
| if (I >= op->getNumOperands()) { | ||
| op->emitError("Index exceeds operand num.\n"); | ||
| return nullptr; | ||
| } | ||
| return createBroadcastOperand(op->getLoc(), b, ty, op->getOperand(I)); | ||
| } | ||
| }; | ||
|
|
||
| struct ConstantOperand { | ||
| template <int64_t I> | ||
| static Value getConst(Operation *op, PatternRewriter &b, TensorType ty) { | ||
| const auto loc = op->getLoc(); | ||
| if (llvm::isa<IntegerType>(ty.getElementType())) { | ||
| return b.create<arith::ConstantOp>( // | ||
| loc, DenseElementsAttr::get(ty, int64_t(I))); | ||
| } else if (llvm::isa<FloatType>(ty.getElementType())) { | ||
| return b.create<arith::ConstantOp>( // | ||
| loc, DenseElementsAttr::get(ty, float(I))); | ||
| } else { | ||
| op->emitError("Not a supported element type for constant.\n"); | ||
| return nullptr; | ||
| } | ||
| } | ||
| }; |
There was a problem hiding this comment.
These methods are stateless. What are they wrapped into a struct for? Is it just for logical grouping expecting more methods?
There was a problem hiding this comment.
Just for logical grouping, will add more, e.g. get constant operand from attrs like onednn_graph.pow
[Transform] Add basic onednn_graph dialect lowering (#61)
Added basic onednn_graph dialect lowering for matmul, relu, add.
Tracking: #28
Depending on: #43
The lowering of elementwise binary ops with constant operand will use named ops, same as the other elementwise ops for consistency. And we will develop and explore its strength/limitation:
We can explore more options in the future if named op cannot comply with our design, e.g.: