1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
|
//===- LoopUnrollAndJam.cpp - Code to perform loop unroll and jam ---------===//
//
// 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 loop unroll and jam. Unroll and jam is a transformation
// that improves locality, in particular, register reuse, while also improving
// operation level parallelism. The example below shows what it does in nearly
// the general case. Loop unroll and jam currently works if the bounds of the
// loops inner to the loop being unroll-jammed do not depend on the latter.
//
// Before After unroll and jam of i by factor 2:
//
// for i, step = 2
// for i S1(i);
// S1; S2(i);
// S2; S1(i+1);
// for j S2(i+1);
// S3; for j
// S4; S3(i, j);
// S5; S4(i, j);
// S6; S3(i+1, j)
// S4(i+1, j)
// S5(i);
// S6(i);
// S5(i+1);
// S6(i+1);
//
// Note: 'if/else' blocks are not jammed. So, if there are loops inside if
// op's, bodies of those loops will not be jammed.
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Affine/Passes.h"
#include "mlir/Dialect/Affine/Analysis/AffineAnalysis.h"
#include "mlir/Dialect/Affine/Analysis/LoopAnalysis.h"
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/Affine/LoopUtils.h"
#include "mlir/Dialect/Func/IR/FuncOps.h"
#include "mlir/IR/AffineExpr.h"
#include "mlir/IR/AffineMap.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/IRMapping.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/Support/CommandLine.h"
#include <optional>
namespace mlir {
namespace affine {
#define GEN_PASS_DEF_AFFINELOOPUNROLLANDJAM
#include "mlir/Dialect/Affine/Passes.h.inc"
} // namespace affine
} // namespace mlir
#define DEBUG_TYPE "affine-loop-unroll-jam"
using namespace mlir;
using namespace mlir::affine;
namespace {
/// Loop unroll jam pass. Currently, this just unroll jams the first
/// outer loop in a Function.
struct LoopUnrollAndJam
: public affine::impl::AffineLoopUnrollAndJamBase<LoopUnrollAndJam> {
explicit LoopUnrollAndJam(
std::optional<unsigned> unrollJamFactor = std::nullopt) {
if (unrollJamFactor)
this->unrollJamFactor = *unrollJamFactor;
}
void runOnOperation() override;
};
} // namespace
std::unique_ptr<OperationPass<func::FuncOp>>
mlir::affine::createLoopUnrollAndJamPass(int unrollJamFactor) {
return std::make_unique<LoopUnrollAndJam>(
unrollJamFactor == -1 ? std::nullopt
: std::optional<unsigned>(unrollJamFactor));
}
void LoopUnrollAndJam::runOnOperation() {
if (getOperation().isExternal())
return;
// Currently, just the outermost loop from the first loop nest is
// unroll-and-jammed by this pass. However, runOnAffineForOp can be called on
// any for operation.
auto &entryBlock = getOperation().front();
if (auto forOp = dyn_cast<AffineForOp>(entryBlock.front()))
(void)loopUnrollJamByFactor(forOp, unrollJamFactor);
}
|
