1 // RUN: mlir-opt %s -split-input-file | FileCheck %s
6 func.func @test0(%arg0 : index, %arg1 : index) {
7 %0 = memref.alloc() : memref<100x100xf32>
8 %1 = memref.alloc() : memref<100x100xf32, affine_map<(d0, d1) -> (d0, d1)>, 2>
9 %2 = memref.alloc() : memref<1xi32>
10 %c0 = arith.constant 0 : index
11 %c64 = arith.constant 64 : index
12 affine.for %i0 = 0 to 10 {
13 affine.for %i1 = 0 to 10 {
14 affine.dma_start %0[%i0, %i1], %1[%i0, %i1], %2[%c0], %c64
15 : memref<100x100xf32>, memref<100x100xf32, 2>, memref<1xi32>
16 affine.dma_wait %2[%c0], %c64 : memref<1xi32>
17 // CHECK: affine.dma_start %{{.*}}[%{{.*}}, %{{.*}}], %{{.*}}[%{{.*}}, %{{.*}}], %{{.*}}[%{{.*}}], %{{.*}} : memref<100x100xf32>, memref<100x100xf32, 2>, memref<1xi32>
18 // CHECK: affine.dma_wait %{{.*}}[%{{.*}}], %{{.*}} : memref<1xi32>
26 // Test with loop IVs and optional stride arguments.
27 func.func @test1(%arg0 : index, %arg1 : index) {
28 %0 = memref.alloc() : memref<100x100xf32>
29 %1 = memref.alloc() : memref<100x100xf32, affine_map<(d0, d1) -> (d0, d1)>, 2>
30 %2 = memref.alloc() : memref<1xi32>
31 %c0 = arith.constant 0 : index
32 %c64 = arith.constant 64 : index
33 %c128 = arith.constant 128 : index
34 %c256 = arith.constant 256 : index
35 affine.for %i0 = 0 to 10 {
36 affine.for %i1 = 0 to 10 {
37 affine.dma_start %0[%i0, %i1], %1[%i0, %i1], %2[%c0], %c64, %c128, %c256
38 : memref<100x100xf32>, memref<100x100xf32, 2>, memref<1xi32>
39 affine.dma_wait %2[%c0], %c64 : memref<1xi32>
40 // CHECK: affine.dma_start %{{.*}}[%{{.*}}, %{{.*}}], %{{.*}}[%{{.*}}, %{{.*}}], %{{.*}}[%{{.*}}], %{{.*}}, %{{.*}}, %{{.*}} : memref<100x100xf32>, memref<100x100xf32, 2>, memref<1xi32>
41 // CHECK: affine.dma_wait %{{.*}}[%{{.*}}], %{{.*}} : memref<1xi32>
49 // Test with loop IVs and symbols (without symbol keyword).
50 func.func @test2(%arg0 : index, %arg1 : index) {
51 %0 = memref.alloc() : memref<100x100xf32>
52 %1 = memref.alloc() : memref<100x100xf32, affine_map<(d0, d1) -> (d0, d1)>, 2>
53 %2 = memref.alloc() : memref<1xi32>
54 %c0 = arith.constant 0 : index
55 %c64 = arith.constant 64 : index
56 affine.for %i0 = 0 to 10 {
57 affine.for %i1 = 0 to 10 {
58 affine.dma_start %0[%i0 + %arg0, %i1], %1[%i0, %i1 + %arg1 + 5],
60 : memref<100x100xf32>, memref<100x100xf32, 2>, memref<1xi32>
61 affine.dma_wait %2[%c0], %c64 : memref<1xi32>
62 // CHECK: affine.dma_start %{{.*}}[%{{.*}} + %{{.*}}, %{{.*}}], %{{.*}}[%{{.*}}, %{{.*}} + %{{.*}} + 5], %{{.*}}[%{{.*}}], %{{.*}} : memref<100x100xf32>, memref<100x100xf32, 2>, memref<1xi32>
63 // CHECK: affine.dma_wait %{{.*}}[%{{.*}}], %{{.*}} : memref<1xi32>
71 // Test with loop IVs and symbols (with symbol keyword).
72 func.func @test3(%arg0 : index, %arg1 : index) {
73 %0 = memref.alloc() : memref<100x100xf32>
74 %1 = memref.alloc() : memref<100x100xf32, affine_map<(d0, d1) -> (d0, d1)>, 2>
75 %2 = memref.alloc() : memref<1xi32>
76 %c0 = arith.constant 0 : index
77 %c64 = arith.constant 64 : index
78 affine.for %i0 = 0 to 10 {
79 affine.for %i1 = 0 to 10 {
80 affine.dma_start %0[%i0 + symbol(%arg0), %i1],
81 %1[%i0, %i1 + symbol(%arg1) + 7],
82 %2[%i0 + %i1 + 11], %c64
83 : memref<100x100xf32>, memref<100x100xf32, 2>, memref<1xi32>
84 affine.dma_wait %2[%c0], %c64 : memref<1xi32>
85 // CHECK: affine.dma_start %{{.*}}[%{{.*}} + symbol(%{{.*}}), %{{.*}}], %{{.*}}[%{{.*}}, %{{.*}} + symbol(%{{.*}}) + 7], %{{.*}}[%{{.*}} + %{{.*}} + 11], %{{.*}} : memref<100x100xf32>, memref<100x100xf32, 2>, memref<1xi32>
86 // CHECK: affine.dma_wait %{{.*}}[%{{.*}}], %{{.*}} : memref<1xi32>
94 // Test with loop IVs, symbols and constants in nested affine expressions.
95 func.func @test4(%arg0 : index, %arg1 : index) {
96 %0 = memref.alloc() : memref<100x100xf32>
97 %1 = memref.alloc() : memref<100x100xf32, 2>
98 %2 = memref.alloc() : memref<1xi32>
99 %c64 = arith.constant 64 : index
100 affine.for %i0 = 0 to 10 {
101 affine.for %i1 = 0 to 10 {
102 affine.dma_start %0[(%i0 + symbol(%arg0)) floordiv 3, %i1],
103 %1[%i0, (%i1 + symbol(%arg1)) mod 9 + 7],
104 %2[%i0 + %i1 + 11], %c64
105 : memref<100x100xf32>, memref<100x100xf32, 2>, memref<1xi32>
106 affine.dma_wait %2[%i0 + %i1 + 11], %c64 : memref<1xi32>
107 // CHECK: affine.dma_start %{{.*}}[(%{{.*}} + symbol(%{{.*}})) floordiv 3, %{{.*}}], %{{.*}}[%{{.*}}, (%{{.*}} + symbol(%{{.*}})) mod 9 + 7], %{{.*}}[%{{.*}} + %{{.*}} + 11], %{{.*}} : memref<100x100xf32>, memref<100x100xf32, 2>, memref<1xi32>
108 // CHECK: affine.dma_wait %{{.*}}[%{{.*}} + %{{.*}} + 11], %{{.*}} : memref<1xi32>