[Mlir-commits] [mlir] [mlir][linalg] unfold projected permutation. (PR #114704)

[Andrzej Warzyński]

================
@@ -0,0 +1,243 @@
+//===- UnfoldProjectedPermutation.cpp - extract projected projections   ---===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+#include "mlir/Dialect/Linalg/Transforms/Transforms.h"
+#include <utility>
+
+#include "mlir/Dialect/Affine/IR/AffineOps.h"
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
+#include <map>
+#include <optional>
+#include <vector>
+
+using namespace mlir;
+using namespace mlir::linalg;
+
+namespace {
+
+/// This file implements pattern to decompose the input operand(s) of a
+/// linalg.generic that has a `transpose`, `broadcast` or a mixture of two,
+/// into explicit transpose and broadcast. Having them folded into the
+/// linalg.generic is a good optimization but sometimes we may want to unwrap
+/// i.e. `unfold` them as explicit transpose and broadcast. This rewrite
+/// pattern helps do it for each input operand. This is useful for instance
+/// when trying to recognize named ops.
+///
+/// The transpose, broadcast, or mixture of both, are expressed in the affine
+/// map of the operand. Technically it is essentially `projected permutation`.
+///
+///  Example
+///
+/// ```mlir
+///
+/// #projection = affine_map<(d0, d1, d2, d3, d4) -> (d2, d3, d1)>
+/// #identity   = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>
+/// ...
+///    %res = linalg.generic
+///       { indexing_maps = [#projection, #identity, #identity],
+///       iterator_types = ["parallel", "parallel", "parallel",
+///                         "parallel", "parallel"]}
+///       ins(%x, %y : tensor<7x8x9xf32>, tensor<5x9x7x8x10xf32>)
+///       outs(%z : tensor<5x9x7x8x10xf32>) {
+///         ^bb0(%in: f32, %in_1: f32, %out: f32):
+///              %div = arith.divf %in, %in_1 : f32
+///              linalg.yield %div : f32
+///    } -> tensor<5x9x7x8x10xf32>
+/// ```
+///
+/// In the above IR operand `%x` map is a projected-permutation. This can be
+/// unfolded as:
+///
+/// ```mlir
+///   ...
+///   %x_trans = linalg.transpose
+///                   ins(%x : tensor<7x8x9xf32>)
+///                   outs(%e1 : tensor<9x7x8xf32>) permutation = [2, 0, 1]
+///   ...
+///   %x_trans_bc = linalg.broadcast
+///                   ins(%x_trans : tensor<9x7x8xf32>)
+///                   outs(%e2 : tensor<5x9x7x8x10xf32>) dimensions = [0, 4]
+///   %2 = linalg.div
+///           ins(%x_trans_bc, %y :
+///                  tensor<5x9x7x8x10xf32>, tensor<5x9x7x8x10xf32>)
+///           outs(%arg2 : tensor<5x9x7x8x10xf32>) -> tensor<5x9x7x8x10xf32>
+///
+/// Note that linalg.generic has been 'specialized' to linalg.div.
+///
+/// To unfold it is more effective to transpose first and then do the broadcast.
+/// However, if transpose is done first, the permutation map needs to be
+/// expressed in terms of reduced dimension (as broadcast hasn't happened yet).
+/// Also, the broadcast dimensions in a linalg.generic come from other operands
+/// (those not broadcasted along that particular dimension). We work this out
+/// by computing the convex-polyhedron shape of the linalg.gneric iteration
----------------
banach-space wrote:

```suggestion
/// by computing the convex-polyhedron shape of the linalg.generic iteration
```

https://github.com/llvm/llvm-project/pull/114704