1 // RUN: mlir-opt %s -split-input-file -test-affine-access-analysis -verify-diagnostics | FileCheck %s
3 // CHECK-LABEL: func @loop_simple
4 func.func @loop_simple(%A : memref<?x?xf32>, %B : memref<?x?x?xf32>) {
5 %c0 = arith.constant 0 : index
6 %M = memref.dim %A, %c0 : memref<?x?xf32>
7 affine.for %i = 0 to %M {
8 affine.for %j = 0 to %M {
9 affine.load %A[%c0, %i] : memref<?x?xf32>
10 // expected-remark@above {{contiguous along loop 0}}
11 // expected-remark@above {{invariant along loop 1}}
12 affine.load %A[%c0, 8 * %i + %j] : memref<?x?xf32>
13 // expected-remark@above {{contiguous along loop 1}}
14 // Note/FIXME: access stride isn't being checked.
15 // expected-remark@-3 {{contiguous along loop 0}}
17 // These are all non-contiguous along both loops. Nothing is emitted.
18 affine.load %A[%i, %c0] : memref<?x?xf32>
19 // expected-remark@above {{invariant along loop 1}}
20 // Note/FIXME: access stride isn't being checked.
21 affine.load %A[%i, 8 * %j] : memref<?x?xf32>
22 // expected-remark@above {{contiguous along loop 1}}
23 affine.load %A[%j, 4 * %i] : memref<?x?xf32>
24 // expected-remark@above {{contiguous along loop 0}}
32 // CHECK-LABEL: func @loop_unsimplified
33 func.func @loop_unsimplified(%A : memref<100xf32>) {
34 affine.for %i = 0 to 100 {
35 affine.load %A[2 * %i - %i - %i] : memref<100xf32>
36 // expected-remark@above {{invariant along loop 0}}
38 %m = affine.apply affine_map<(d0) -> (-2 * d0)>(%i)
39 %n = affine.apply affine_map<(d0) -> (2 * d0)>(%i)
40 affine.load %A[(%m + %n) floordiv 2] : memref<100xf32>
41 // expected-remark@above {{invariant along loop 0}}
48 #map = affine_map<(d0) -> (d0 * 16)>
49 #map1 = affine_map<(d0) -> (d0 * 16 + 16)>
50 #map2 = affine_map<(d0) -> (d0)>
51 #map3 = affine_map<(d0) -> (d0 + 1)>
53 func.func @tiled(%arg0: memref<*xf32>) {
54 %alloc = memref.alloc() {alignment = 64 : i64} : memref<1x224x224x64xf32>
55 %cast = memref.cast %arg0 : memref<*xf32> to memref<64xf32>
56 affine.for %arg1 = 0 to 4 {
57 affine.for %arg2 = 0 to 224 {
58 affine.for %arg3 = 0 to 14 {
59 %alloc_0 = memref.alloc() : memref<1x16x1x16xf32>
60 affine.for %arg4 = #map(%arg1) to #map1(%arg1) {
61 affine.for %arg5 = #map(%arg3) to #map1(%arg3) {
62 // TODO: here and below, the access isn't really invariant
63 // along tile-space IVs where the intra-tile IVs' bounds
65 %0 = affine.load %cast[%arg4] : memref<64xf32>
66 // expected-remark@above {{contiguous along loop 3}}
67 // expected-remark@above {{invariant along loop 0}}
68 // expected-remark@above {{invariant along loop 1}}
69 // expected-remark@above {{invariant along loop 2}}
70 // expected-remark@above {{invariant along loop 4}}
71 affine.store %0, %alloc_0[0, %arg1 * -16 + %arg4, 0, %arg3 * -16 + %arg5] : memref<1x16x1x16xf32>
72 // expected-remark@above {{contiguous along loop 4}}
73 // expected-remark@above {{contiguous along loop 2}}
74 // expected-remark@above {{invariant along loop 1}}
77 affine.for %arg4 = #map(%arg1) to #map1(%arg1) {
78 affine.for %arg5 = #map2(%arg2) to #map3(%arg2) {
79 affine.for %arg6 = #map(%arg3) to #map1(%arg3) {
80 %0 = affine.load %alloc_0[0, %arg1 * -16 + %arg4, -%arg2 + %arg5, %arg3 * -16 + %arg6] : memref<1x16x1x16xf32>
81 // expected-remark@above {{contiguous along loop 5}}
82 // expected-remark@above {{contiguous along loop 2}}
83 affine.store %0, %alloc[0, %arg5, %arg6, %arg4] : memref<1x224x224x64xf32>
84 // expected-remark@above {{contiguous along loop 3}}
85 // expected-remark@above {{invariant along loop 0}}
86 // expected-remark@above {{invariant along loop 1}}
87 // expected-remark@above {{invariant along loop 2}}
91 memref.dealloc %alloc_0 : memref<1x16x1x16xf32>