[Mlir-commits] [mlir] [mlir][Vector] Add `vector.shuffle` tree transformation (PR #145740)

[Diego Caballero]

================
@@ -0,0 +1,692 @@
+//===- VectorShuffleTreeBuilder.cpp ----- Vector shuffle tree builder -----===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements pattern rewrites to lower sequences of
+// `vector.to_elements` and `vector.from_elements` operations into a tree of
+// `vector.shuffle` operations.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Vector/IR/VectorOps.h"
+#include "mlir/Dialect/Vector/Transforms/LoweringPatterns.h"
+#include "mlir/Dialect/Vector/Transforms/Passes.h"
+#include "mlir/Rewrite/FrozenRewritePatternSet.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/MathExtras.h"
+#include "llvm/Support/raw_ostream.h"
+
+namespace mlir {
+namespace vector {
+
+#define GEN_PASS_DEF_LOWERVECTORTOFROMELEMENTSTOSHUFFLETREE
+#include "mlir/Dialect/Vector/Transforms/Passes.h.inc"
+
+} // namespace vector
+} // namespace mlir
+
+#define DEBUG_TYPE "lower-vector-to-from-elements-to-shuffle-tree"
+
+using namespace mlir;
+using namespace mlir::vector;
+
+namespace {
+
+// Indentation unit for debug output formatting.
+constexpr unsigned kIndScale = 2;
+
+/// Represents a closed interval of elements (e.g., [0, 7] = 8 elements).
+using Interval = std::pair<unsigned, unsigned>;
+// Sentinel value for uninitialized intervals.
+constexpr unsigned kMaxUnsigned = std::numeric_limits<unsigned>::max();
+
+/// The VectorShuffleTreeBuilder builds a balanced binary tree of
+/// `vector.shuffle` operations from one or more `vector.to_elements`
+/// operations feeding a single `vector.from_elements` operation.
+///
+/// The implementation generates hardware-agnostic `vector.shuffle` operations
+/// that minimize both the number of shuffle operations and the length of
+/// intermediate vectors (to the extent possible). The tree has the
+/// following properties:
+///
+///   1. Vectors are shuffled in pairs by order of appearance in
+///      the `vector.from_elements` operand list.
+///   2. Each input vector to each level is used only once.
+///   3. The number of levels in the tree is:
+///        ceil(log2(# `vector.to_elements` ops)).
+///   4. Vectors at each level of the tree have the same vector length.
----------------
dcaballe wrote:

The current implementation bails out, although it should be easy to support...

https://github.com/llvm/llvm-project/pull/145740