[mlir][linalg] Add folder for linalg.index

[kuhar]

We know that the index of unit dims is always 0.

[llvmbot]

@llvm/pr-subscribers-mlir-linalg

@llvm/pr-subscribers-mlir

Author: Jakub Kuderski (kuhar)

Changes

We know that the index of unit dims is always 0.

---

Full diff: https://github.com/llvm/llvm-project/pull/136640.diff

3 Files Affected:

• (modified) mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td (+1)
• (modified) mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp (+29)
• (modified) mlir/test/Dialect/Linalg/canonicalize.mlir (+80)

diff --git a/mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td b/mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td index f8df828f74851..1b48bf5fcb237 100644 --- a/mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td +++ b/mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td @@ -88,6 +88,7 @@ def Linalg_IndexOp : Linalg_Op<"index", [Pure]>, let assemblyFormat = [{ $dim attr-dict `:` type($result) }]; let hasVerifier = 1; + let hasFolder = 1; } def Linalg_SoftmaxOp : Linalg_Op<"softmax", diff --git a/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp b/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp index 6c680498af2ad..a3787f101afa3 100644 --- a/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp +++ b/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp @@ -2283,6 +2283,35 @@ LogicalResult IndexOp::verify() { return success(); } +OpFoldResult IndexOp::fold(FoldAdaptor adaptor) { + auto linalgOp = cast<LinalgOp>((*this)->getParentOp()); + int64_t flatDimPos = + cast<AffineDimExpr>(linalgOp.getShapesToLoopsMap().getResult(getDim())) + .getPosition(); + + // Find the flat dimension position among the operands. + int64_t flatPosOffset = 0; + for (Value operand : linalgOp->getOperands()) { + assert(flatDimPos >= flatPosOffset && "invalid position"); + auto shapedType = dyn_cast<ShapedType>(operand.getType()); + if (!shapedType) + break; + + int64_t rank = shapedType.getRank(); + if (flatDimPos < flatPosOffset + rank) { + // Found the dimension within this shape. Now we can either fold if the + // dim size is 1, or bail out otherwise. + int64_t pos = flatDimPos - flatPosOffset; + if (shapedType.getDimSize(pos) != 1) + break; + + return IntegerAttr::get(IndexType::get(getContext()), 0); + } + flatPosOffset += rank; + } + return OpFoldResult{}; +} + /////// Operations corresponding to library calls defined with Tablegen //////// #include "mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yamlgen.cpp.inc" diff --git a/mlir/test/Dialect/Linalg/canonicalize.mlir b/mlir/test/Dialect/Linalg/canonicalize.mlir index 86cb8f58abe02..3daf221f4402d 100644 --- a/mlir/test/Dialect/Linalg/canonicalize.mlir +++ b/mlir/test/Dialect/Linalg/canonicalize.mlir @@ -305,6 +305,86 @@ func.func @self_copy(%arg0 : memref<2x3x?x4xf32>) { } // ----- + +// CHECK: func @fold_linalg_index_tensor_static +func.func @fold_linalg_index_tensor_static(%0: tensor<4x16xi32>, %1: tensor<1x16xi32>, + %2: tensor<4x1xi32>) -> tensor<4x1xi32> { +// CHECK-NEXT: linalg.generic +// CHECK: %[[IDX_0:.+]] = linalg.index 0 : index +// CHECK-NOT: linalg.index 1 +// CHECK: %[[IDX_2:.+]] = linalg.index 2 : index +// CHECK: %[[ADD:.+]] = arith.addi %[[IDX_0]], %[[IDX_2]] +// CHECK: %[[CAST:.+]] = arith.index_cast %[[ADD]] +// CHECK: linalg.yield %[[CAST]] + %res = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, + affine_map<(d0, d1, d2) -> (d1, d2)>, + affine_map<(d0, d1, d2) -> (d0, d1)>], + iterator_types = ["parallel", "parallel", "reduction"]} + ins(%0, %1 : tensor<4x16xi32>, tensor<1x16xi32>) + outs(%2 : tensor<4x1xi32>) { + ^bb0(%lhs: i32, %rhs: i32, %out: i32): + %idx0 = linalg.index 0 : index + %idx1 = linalg.index 1 : index + %idx2 = linalg.index 2 : index + %add0 = arith.addi %idx0, %idx1 : index + %add1 = arith.addi %add0, %idx2 : index + %int = arith.index_cast %add1 : index to i32 + linalg.yield %int : i32 + } -> tensor<4x1xi32> + return %res : tensor<4x1xi32> +} + +// ----- + +// CHECK: func @fold_linalg_index_tensor_dynamic +func.func @fold_linalg_index_tensor_dynamic(%0: tensor<?x1xi32>, + %1: tensor<?x1xi32>) -> tensor<?x1xi32> { +// CHECK-NEXT: linalg.generic +// CHECK: %[[IDX_0:.+]] = linalg.index 0 : index +// CHECK-NOT: linalg.index 1 +// CHECK: %[[CAST:.+]] = arith.index_cast %[[IDX_0]] +// CHECK: linalg.yield %[[CAST]] + %res = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, + affine_map<(d0, d1) -> (d1, d1)>], + iterator_types = ["parallel", "parallel"]} + ins(%0 : tensor<?x1xi32>) + outs(%1 : tensor<?x1xi32>) { + ^bb0(%lhs: i32, %out: i32): + %idx0 = linalg.index 0 : index + %idx1 = linalg.index 1 : index + %add = arith.addi %idx0, %idx1 : index + %int = arith.index_cast %add : index to i32 + linalg.yield %int : i32 + } -> tensor<?x1xi32> + return %res : tensor<?x1xi32> +} + +// ----- + +// CHECK: func @fold_linalg_index_memref +func.func @fold_linalg_index_memref(%0: memref<1x?xi32>, %1: memref<1x?xi32>) { +// CHECK-NEXT: linalg.generic +// CHECK-NOT: linalg.index 0 +// CHECK: %[[IDX_1:.+]] = linalg.index 1 : index +// CHECK: %[[CAST:.+]] = arith.index_cast %[[IDX_1]] +// CHECK: linalg.yield %[[CAST]] + linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, + affine_map<(d0, d1) -> (d1, d1)>], + iterator_types = ["parallel", "parallel"]} + ins(%0 : memref<1x?xi32>) + outs(%1 : memref<1x?xi32>) { + ^bb0(%lhs: i32, %out: i32): + %idx0 = linalg.index 0 : index + %idx1 = linalg.index 1 : index + %add = arith.addi %idx0, %idx1 : index + %int = arith.index_cast %add : index to i32 + linalg.yield %int : i32 + } + return +} + +// ----- + // CHECK-LABEL: func @fold_fill_reshape() func.func @fold_fill_reshape() -> tensor<6x4xf32> { %zero = arith.constant 0.0 : f32 

[banach-space]

Makes sense, thanks!

I've left some nits/side-comments. Feel free to ignore.