107 lines · plain
1// RUN: mlir-opt %s -allow-unregistered-dialect -test-multi-buffering=multiplier=5 -cse -split-input-file | FileCheck %s2 3// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0) -> (((d0 - 1) floordiv 3) mod 5)>4 5// CHECK-LABEL: func @multi_buffer6func.func @multi_buffer(%a: memref<1024x1024xf32>) {7// CHECK-DAG: %[[A:.*]] = memref.alloc() {someAttribute} : memref<5x4x128xf32>8// CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index9// CHECK-DAG: %[[C3:.*]] = arith.constant 3 : index10 %0 = memref.alloc() {someAttribute} : memref<4x128xf32>11 %c1024 = arith.constant 1024 : index12 %c1 = arith.constant 1 : index13 %c3 = arith.constant 3 : index14// CHECK: scf.for %[[IV:.*]] = %[[C1]]15 scf.for %arg2 = %c1 to %c1024 step %c3 {16// CHECK: %[[I:.*]] = affine.apply #[[$MAP1]](%[[IV]])17// CHECK: %[[SV:.*]] = memref.subview %[[A]][%[[I]], 0, 0] [1, 4, 128] [1, 1, 1] : memref<5x4x128xf32> to memref<4x128xf32, strided<[128, 1], offset: ?>>18 %1 = memref.subview %a[%arg2, 0] [4, 128] [1, 1] :19 memref<1024x1024xf32> to memref<4x128xf32, affine_map<(d0, d1)[s0] -> (d0 * 1024 + s0 + d1)>>20// CHECK: memref.copy %{{.*}}, %[[SV]] : memref<4x128xf32, #{{.*}}> to memref<4x128xf32, strided<[128, 1], offset: ?>>21 memref.copy %1, %0 : memref<4x128xf32, affine_map<(d0, d1)[s0] -> (d0 * 1024 + s0 + d1)>> to memref<4x128xf32>22// CHECK: "some_use"(%[[SV]]) : (memref<4x128xf32, strided{{.*}}>) -> ()23 "some_use"(%0) : (memref<4x128xf32>) -> ()24// CHECK: "some_use"(%[[SV]]) : (memref<4x128xf32, strided{{.*}}>) -> ()25 "some_use"(%0) : (memref<4x128xf32>) -> ()26 }27 return28}29 30// -----31 32// CHECK-LABEL: func @multi_buffer_affine33func.func @multi_buffer_affine(%a: memref<1024x1024xf32>) {34// CHECK-DAG: %[[A:.*]] = memref.alloc() : memref<5x4x128xf32>35 %0 = memref.alloc() : memref<4x128xf32>36 %c1024 = arith.constant 1024 : index37 %c1 = arith.constant 1 : index38 %c3 = arith.constant 3 : index39// CHECK: affine.for %[[IV:.*]] = 140 affine.for %arg2 = 1 to 1024 step 3 {41// CHECK: %[[I:.*]] = affine.apply #[[$MAP1]](%[[IV]])42// CHECK: %[[SV:.*]] = memref.subview %[[A]][%[[I]], 0, 0] [1, 4, 128] [1, 1, 1] : memref<5x4x128xf32> to memref<4x128xf32, strided<[128, 1], offset: ?>>43 %1 = memref.subview %a[%arg2, 0] [4, 128] [1, 1] :44 memref<1024x1024xf32> to memref<4x128xf32, affine_map<(d0, d1)[s0] -> (d0 * 1024 + s0 + d1)>>45// CHECK: memref.copy %{{.*}}, %[[SV]] : memref<4x128xf32, #{{.*}}> to memref<4x128xf32, strided<[128, 1], offset: ?>>46 memref.copy %1, %0 : memref<4x128xf32, affine_map<(d0, d1)[s0] -> (d0 * 1024 + s0 + d1)>> to memref<4x128xf32>47// CHECK: "some_use"(%[[SV]]) : (memref<4x128xf32, strided{{.*}}>) -> ()48 "some_use"(%0) : (memref<4x128xf32>) -> ()49// CHECK: "some_use"(%[[SV]]) : (memref<4x128xf32, strided{{.*}}>) -> ()50 "some_use"(%0) : (memref<4x128xf32>) -> ()51 }52 return53}54 55// -----56 57// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0) -> (((d0 - 1) floordiv 3) mod 5)>58 59// CHECK-LABEL: func @multi_buffer_subview_use60func.func @multi_buffer_subview_use(%a: memref<1024x1024xf32>) {61// CHECK-DAG: %[[A:.*]] = memref.alloc() : memref<5x4x128xf32>62// CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index63// CHECK-DAG: %[[C3:.*]] = arith.constant 3 : index64 %0 = memref.alloc() : memref<4x128xf32>65 %c1024 = arith.constant 1024 : index66 %c1 = arith.constant 1 : index67 %c3 = arith.constant 3 : index68// CHECK: scf.for %[[IV:.*]] = %[[C1]]69 scf.for %arg2 = %c1 to %c1024 step %c3 {70// CHECK: %[[I:.*]] = affine.apply #[[$MAP1]](%[[IV]])71// CHECK: %[[SV:.*]] = memref.subview %[[A]][%[[I]], 0, 0] [1, 4, 128] [1, 1, 1] : memref<5x4x128xf32> to memref<4x128xf32, strided<[128, 1], offset: ?>>72 %1 = memref.subview %a[%arg2, 0] [4, 128] [1, 1] :73 memref<1024x1024xf32> to memref<4x128xf32, affine_map<(d0, d1)[s0] -> (d0 * 1024 + s0 + d1)>>74// CHECK: memref.copy %{{.*}}, %[[SV]] : memref<4x128xf32, #{{.*}}> to memref<4x128xf32, strided<[128, 1], offset: ?>>75 memref.copy %1, %0 : memref<4x128xf32, affine_map<(d0, d1)[s0] -> (d0 * 1024 + s0 + d1)>> to memref<4x128xf32>76// CHECK: %[[SV1:.*]] = memref.subview %[[SV]][0, 1] [4, 127] [1, 1] : memref<4x128xf32, strided<[128, 1], offset: ?>> to memref<4x127xf32, strided<[128, 1], offset: ?>>77 %s = memref.subview %0[0, 1] [4, 127] [1, 1] :78 memref<4x128xf32> to memref<4x127xf32, affine_map<(d0, d1) -> (d0 * 128 + d1 + 1)>>79// CHECK: "some_use"(%[[SV1]]) : (memref<4x127xf32, strided<[128, 1], offset: ?>>) -> ()80 "some_use"(%s) : (memref<4x127xf32, affine_map<(d0, d1) -> (d0 * 128 + d1 + 1)>>) -> ()81// CHECK: "some_use"(%[[SV]]) : (memref<4x128xf32, strided<[128, 1], offset: ?>>) -> ()82 "some_use"(%0) : (memref<4x128xf32>) -> ()83 }84 return85}86 87// -----88 89// CHECK-LABEL: func @multi_buffer_negative90func.func @multi_buffer_negative(%a: memref<1024x1024xf32>) {91// CHECK-NOT: %{{.*}} = memref.alloc() : memref<5x4x128xf32>92// CHECK: %{{.*}} = memref.alloc() : memref<4x128xf32>93 %0 = memref.alloc() : memref<4x128xf32>94 %c1024 = arith.constant 1024 : index95 %c0 = arith.constant 0 : index96 %c3 = arith.constant 3 : index97 scf.for %arg2 = %c0 to %c1024 step %c3 {98 "blocking_use"(%0) : (memref<4x128xf32>) -> ()99 %1 = memref.subview %a[%arg2, 0] [4, 128] [1, 1] :100 memref<1024x1024xf32> to memref<4x128xf32, affine_map<(d0, d1)[s0] -> (d0 * 1024 + s0 + d1)>>101 memref.copy %1, %0 : memref<4x128xf32, affine_map<(d0, d1)[s0] -> (d0 * 1024 + s0 + d1)>> to memref<4x128xf32>102 "some_use"(%0) : (memref<4x128xf32>) -> ()103 }104 return105}106 107