326 lines · plain
1// RUN: mlir-opt %s -affine-loop-normalize -split-input-file | FileCheck %s2// RUN: mlir-opt %s -affine-loop-normalize='promote-single-iter=1' -split-input-file | FileCheck %s --check-prefix=PROMOTE-SINGLE-ITER3 4// Normalize steps to 1 and lower bounds to 0.5 6// CHECK-DAG: [[$MAP0:#map[0-9]*]] = affine_map<(d0) -> (d0 * 3)>7// CHECK-DAG: [[$MAP1:#map[0-9]*]] = affine_map<(d0) -> (d0 * 2 + 1)>8// CHECK-DAG: [[$MAP2:#map[0-9]*]] = affine_map<(d0, d1) -> (d0 + d1)>9 10// CHECK-LABEL: func @normalize_parallel()11func.func @normalize_parallel() {12 %cst = arith.constant 1.0 : f3213 %0 = memref.alloc() : memref<2x4xf32>14 // CHECK: affine.parallel (%[[i0:.*]], %[[j0:.*]]) = (0, 0) to (4, 2)15 affine.parallel (%i, %j) = (0, 1) to (10, 5) step (3, 2) {16 // CHECK: %[[i1:.*]] = affine.apply [[$MAP0]](%[[i0]])17 // CHECK: %[[j1:.*]] = affine.apply [[$MAP1]](%[[j0]])18 // CHECK: affine.parallel (%[[k0:.*]]) = (0) to (%[[j1]] - %[[i1]])19 affine.parallel (%k) = (%i) to (%j) {20 // CHECK: %[[k1:.*]] = affine.apply [[$MAP2]](%[[i1]], %[[k0]])21 // CHECK: affine.store %{{.*}}, %{{.*}}[%[[i1]], %[[k1]]] : memref<2x4xf32>22 affine.store %cst, %0[%i, %k] : memref<2x4xf32>23 }24 }25 return26}27 28// -----29 30// CHECK-LABEL: func @relative_bounds31func.func @relative_bounds(%arg: index) {32 // CHECK: affine.for %{{.*}} = 0 to 433 affine.for %i = affine_map<(d0) -> (d0)>(%arg) to affine_map<(d0) -> (d0 + 4)>(%arg) {34 }35 return36}37 38// -----39 40// Check that single iteration loop is removed and its body is promoted to the41// parent block.42 43// CHECK-LABEL: func @promote_single_iter_loop44// PROMOTE-SINGLE-ITER-LABEL: func @promote_single_iter_loop45func.func @promote_single_iter_loop(%in: memref<1xf32>, %out: memref<1xf32>) {46 affine.for %i = 0 to 1 {47 %1 = affine.load %in[%i] : memref<1xf32>48 affine.store %1, %out[%i] : memref<1xf32>49 }50 return51}52 53// PROMOTE-SINGLE-ITER-NEXT: arith.constant54// PROMOTE-SINGLE-ITER-NEXT: affine.load55// PROMOTE-SINGLE-ITER-NEXT: affine.store56// PROMOTE-SINGLE-ITER-NEXT: return57 58// -----59 60// CHECK-DAG: [[$IV0:#map[0-9]*]] = affine_map<(d0) -> (d0 * 2 + 2)>61// CHECK-DAG: [[$IV1:#map[0-9]*]] = affine_map<(d0) -> (d0 * 3)>62 63// CHECK-LABEL: func @simple_loop_nest()64// CHECK-NEXT: affine.for %[[I:.*]] = 0 to 15 {65// CHECK-NEXT: %[[IIV:.*]] = affine.apply [[$IV0]](%[[I]])66// CHECK-NEXT: affine.for %[[II:.*]] = 0 to 11 {67// CHECK-NEXT: %[[IIIV:.*]] = affine.apply [[$IV1]](%[[II]])68// CHECK-NEXT: "test.foo"(%[[IIV]], %[[IIIV]])69// CHECK-NEXT: }70// CHECK-NEXT: }71// CHECK-NEXT: return72// CHECK-NEXT: }73func.func @simple_loop_nest(){74 affine.for %i0 = 2 to 32 step 2 {75 affine.for %i1 = 0 to 32 step 3 {76 "test.foo"(%i0, %i1) : (index, index) -> ()77 }78 }79 return80}81 82// -----83 84// CHECK-DAG: [[$IV00:#map[0-9]*]] = affine_map<(d0) -> (d0 * 32 + 2)>85// CHECK-DAG: [[$IV11:#map[0-9]*]] = affine_map<(d0) -> (d0 * 2)>86// CHECK-DAG: [[$UB00:#map[0-9]*]] = affine_map<()[s0] -> ((s0 - 2) ceildiv 32)>87// CHECK-DAG: [[$UB11:#map[0-9]*]] = affine_map<()[s0] -> (s0 ceildiv 2)>88 89// CHECK-LABEL: func @loop_with_unknown_upper_bound90// CHECK-SAME: (%[[ARG0:.*]]: memref<?x?xf32>, %[[ARG1:.*]]: index)91// CHECK-NEXT: arith.constant 0 : index92// CHECK-NEXT: %[[DIM:.*]] = memref.dim %arg0, %c0 : memref<?x?xf32>93// CHECK-NEXT: affine.for %[[I:.*]] = 0 to [[$UB00]]()[%[[DIM]]] {94// CHECK-NEXT: %[[IIV:.*]] = affine.apply [[$IV00]](%[[I]])95// CHECK-NEXT: affine.for %[[II:.*]] = 0 to [[$UB11]]()[%[[ARG1]]] {96// CHECK-NEXT: %[[IIIV:.*]] = affine.apply [[$IV11]](%[[II]])97// CHECK-NEXT: "test.foo"(%[[IIV]], %[[IIIV]])98// CHECK-NEXT: }99// CHECK-NEXT: }100// CHECK-NEXT: return101// CHECK-NEXT: }102func.func @loop_with_unknown_upper_bound(%arg0: memref<?x?xf32>, %arg1: index) {103 %c0 = arith.constant 0 : index104 %0 = memref.dim %arg0, %c0 : memref<?x?xf32>105 affine.for %i0 = 2 to %0 step 32 {106 affine.for %i1 = 0 to %arg1 step 2 {107 "test.foo"(%i0, %i1) : (index, index) -> ()108 }109 }110 return111}112 113// -----114 115// CHECK-DAG: [[$OUTERIV:#map[0-9]*]] = affine_map<(d0) -> (d0 * 32 + 2)>116// CHECK-DAG: [[$INNERIV:#map[0-9]*]] = affine_map<(d0) -> (d0 + 2)>117// CHECK-DAG: [[$OUTERUB:#map[0-9]*]] = affine_map<()[s0] -> ((s0 - 2) ceildiv 32)>118// CHECK-DAG: [[$INNERUB:#map[0-9]*]] = affine_map<()[s0] -> (s0 - 2, 510)>119 120// CHECK-LABEL: func @loop_with_multiple_upper_bounds121// CHECK-SAME: (%[[ARG0:.*]]: memref<?x?xf32>, %[[ARG1:.*]]: index)122// CHECK-NEXT: arith.constant 0 : index123// CHECK-NEXT: %[[DIM:.*]] = memref.dim %arg0, %c0 : memref<?x?xf32>124// CHECK-NEXT: affine.for %[[I:.*]] = 0 to [[$OUTERUB]]()[%[[DIM]]] {125// CHECK-NEXT: %[[IIV:.*]] = affine.apply [[$OUTERIV]](%[[I]])126// CHECK-NEXT: affine.for %[[II:.*]] = 0 to min [[$INNERUB]]()[%[[ARG1]]] {127// CHECK-NEXT: %[[IIIV:.*]] = affine.apply [[$INNERIV]](%[[II]])128// CHECK-NEXT: "test.foo"(%[[IIV]], %[[IIIV]])129// CHECK-NEXT: }130// CHECK-NEXT: }131// CHECK-NEXT: return132// CHECK-NEXT: }133func.func @loop_with_multiple_upper_bounds(%arg0: memref<?x?xf32>, %arg1 : index) {134 %c0 = arith.constant 0 : index135 %0 = memref.dim %arg0, %c0 : memref<?x?xf32>136 affine.for %i0 = 2 to %0 step 32{137 affine.for %i1 = 2 to min affine_map<(d0)[] -> (d0, 512)>(%arg1) {138 "test.foo"(%i0, %i1) : (index, index) -> ()139 }140 }141 return142}143 144// -----145 146// CHECK-DAG: [[$INTERUB:#map[0-9]*]] = affine_map<()[s0] -> (s0 ceildiv 32)>147// CHECK-DAG: [[$INTERIV:#map[0-9]*]] = affine_map<(d0) -> (d0 * 32)>148// CHECK-DAG: [[$INTRAUB:#map[0-9]*]] = affine_map<(d0)[s0] -> (32, -d0 + s0)>149// CHECK-DAG: [[$INTRAIV:#map[0-9]*]] = affine_map<(d0, d1) -> (d0 + d1)>150 151// CHECK-LABEL: func @tiled_matmul152// CHECK-SAME: (%[[ARG0:.*]]: memref<1024x1024xf32>, %[[ARG1:.*]]: memref<1024x1024xf32>, %[[ARG2:.*]]: memref<1024x1024xf32>)153// CHECK-NEXT: arith.constant 0 : index154// CHECK-NEXT: arith.constant 1 : index155// CHECK-NEXT: %[[DIM0:.*]] = memref.dim %[[ARG0]], %{{.*}}156// CHECK-NEXT: %[[DIM1:.*]] = memref.dim %[[ARG1]], %{{.*}}157// CHECK-NEXT: %[[DIM2:.*]] = memref.dim %[[ARG0]], %{{.*}}158// CHECK-NEXT: affine.for %[[I:.*]] = 0 to [[$INTERUB]]()[%[[DIM0]]] {159// CHECK-NEXT: %[[IIV:.*]] = affine.apply [[$INTERIV]](%[[I]])160// CHECK-NEXT: affine.for %[[J:.*]] = 0 to [[$INTERUB]]()[%[[DIM1]]] {161// CHECK-NEXT: %[[JIV:.*]] = affine.apply [[$INTERIV]](%[[J]])162// CHECK-NEXT: affine.for %[[K:.*]] = 0 to [[$INTERUB]]()[%[[DIM2]]] {163// CHECK-NEXT: %[[KIV:.*]] = affine.apply [[$INTERIV]](%[[K]])164// CHECK-NEXT: affine.for %[[II:.*]] = 0 to min [[$INTRAUB]](%[[IIV]])[%[[DIM0]]] {165// CHECK-NEXT: %[[IIIV:.*]] = affine.apply [[$INTRAIV]](%[[IIV]], %[[II]])166// CHECK-NEXT: affine.for %[[JJ:.*]] = 0 to min [[$INTRAUB]](%[[JIV]])[%[[DIM1]]] {167// CHECK-NEXT: %[[JJIV:.*]] = affine.apply [[$INTRAIV]](%[[JIV]], %[[JJ]])168// CHECK-NEXT: affine.for %[[KK:.*]] = 0 to min [[$INTRAUB]](%[[KIV]])[%[[DIM2]]] {169// CHECK-NEXT: %[[KKIV:.*]] = affine.apply [[$INTRAIV]](%[[KIV]], %[[KK]])170// CHECK-NEXT: affine.load %[[ARG0]][%[[IIIV]], %[[KKIV]]] : memref<1024x1024xf32>171// CHECK-NEXT: affine.load %[[ARG1]][%[[KKIV]], %[[JJIV]]] : memref<1024x1024xf32>172// CHECK-NEXT: affine.load %[[ARG2]][%[[IIIV]], %[[JJIV]]] : memref<1024x1024xf32>173// CHECK-NEXT: arith.mulf174// CHECK-NEXT: arith.addf175// CHECK-NEXT: affine.store %{{.*}}, %[[ARG2]]{{.*}} : memref<1024x1024xf32>176// CHECK-NEXT: }177// CHECK-NEXT: }178// CHECK-NEXT: }179// CHECK-NEXT: }180// CHECK-NEXT: }181// CHECK-NEXT: }182// CHECK-NEXT: return183// CHECK-NEXT: }184#map0 = affine_map<(d0, d1) -> (d0, d1)>185#map1 = affine_map<(d0) -> (d0)>186#map2 = affine_map<(d0)[s0] -> (d0 + 32, s0)>187#map3 = affine_map<() -> (0)>188#map4 = affine_map<()[s0] -> (s0)>189 190func.func @tiled_matmul(%0: memref<1024x1024xf32>, %1: memref<1024x1024xf32>, %2: memref<1024x1024xf32>) {191 %c0 = arith.constant 0 : index192 %c1 = arith.constant 1 : index193 %3 = memref.dim %0, %c0 : memref<1024x1024xf32>194 %4 = memref.dim %1, %c1 : memref<1024x1024xf32>195 %5 = memref.dim %0, %c1 : memref<1024x1024xf32>196 affine.for %arg0 = 0 to %3 step 32 {197 affine.for %arg1 = 0 to %4 step 32 {198 affine.for %arg2 = 0 to %5 step 32 {199 affine.for %arg3 = #map1(%arg0) to min #map2(%arg0)[%3] {200 affine.for %arg4 = #map1(%arg1) to min #map2(%arg1)[%4] {201 affine.for %arg5 = #map1(%arg2) to min #map2(%arg2)[%5] {202 %6 = affine.load %0[%arg3, %arg5] : memref<1024x1024xf32>203 %7 = affine.load %1[%arg5, %arg4] : memref<1024x1024xf32>204 %8 = affine.load %2[%arg3, %arg4] : memref<1024x1024xf32>205 %9 = arith.mulf %6, %7 : f32206 %10 = arith.addf %8, %9 : f32207 affine.store %10, %2[%arg3, %arg4] : memref<1024x1024xf32>208 }209 }210 }211 }212 }213 }214 return215}216 217// -----218 219// CHECK-LABEL: func @constant_lower_bound220func.func @constant_lower_bound() {221 %c0 = arith.constant 0 : index222 %c1 = arith.constant 1 : index223 scf.for %j = %c0 to %c1 step %c1 {224 // CHECK: affine.for %[[ARG0:.*]] =225 affine.for %i = %c0 to %c1 {226 // CHECK-NEXT: affine.apply #map{{.*}}(%[[ARG0]])227 }228 }229 return230}231 232// -----233 234// CHECK-DAG: [[$UB_MAP:#map[0-9]*]] = affine_map<()[s0] -> (s0 ceildiv 4)>235// CHECK-DAG: [[$IV_MAP:#map[0-9]*]] = affine_map<(d0) -> (d0 * 4)>236 237// CHECK-LABEL: func @upper_bound_by_symbol238func.func @upper_bound_by_symbol(%arg0: index, %arg1: index) {239 // CHECK: affine.for %[[ARG0:.*]] = 0 to [[$UB_MAP]]()[%arg{{.*}}] {240 affine.for %i = 0 to affine_map<()[s0, s1] -> (s0)>()[%arg0, %arg1] step 4 {241 // CHECK-NEXT: %[[IV:.*]] = affine.apply [[$IV_MAP]](%[[ARG0]])242 // CHECK-NEXT: "test.foo"(%[[IV]]) : (index) -> ()243 "test.foo"(%i) : (index) -> ()244 }245 return246}247 248// -----249 250// CHECK-DAG: [[$UB_MAP:#map[0-9]*]] = affine_map<()[s0] -> ((-s0 + 10) ceildiv 4)>251// CHECK-DAG: [[$IV_MAP:#map[0-9]*]] = affine_map<(d0)[s0] -> (d0 * 4 + s0)>252 253// CHECK-LABEL: func @lower_bound_by_symbol254func.func @lower_bound_by_symbol(%arg0: index, %arg1: index) {255 // CHECK: affine.for %[[ARG0:.*]] = 0 to [[$UB_MAP]]()[%arg{{.*}}] {256 affine.for %i = affine_map<()[s0, s1] -> (s0)>()[%arg0, %arg1] to 10 step 4 {257 // CHECK-NEXT: %[[IV:.*]] = affine.apply [[$IV_MAP]](%[[ARG0]])[%arg{{.*}}]258 // CHECK-NEXT: "test.foo"(%[[IV]]) : (index) -> ()259 "test.foo"(%i) : (index) -> ()260 }261 return262}263 264// -----265 266// CHECK-DAG: [[$UB_MAP:#map[0-9]*]] = affine_map<()[s0] -> (s0 ceildiv 4)>267// CHECK-DAG: [[$IV_MAP:#map[0-9]*]] = affine_map<(d0) -> (d0 * 4)>268 269// CHECK-LABEL: func @upper_bound_by_dim270func.func @upper_bound_by_dim(%arg0: index, %arg1: index) {271 // CHECK: affine.for %[[ARG0:.*]] = 0 to [[$UB_MAP]]()[%arg{{.*}}] {272 affine.for %i = 0 to affine_map<(d0, d1) -> (d0)>(%arg0, %arg1) step 4 {273 // CHECK-NEXT: %[[IV:.*]] = affine.apply [[$IV_MAP]](%[[ARG0]])274 // CHECK-NEXT: "test.foo"(%[[IV]]) : (index) -> ()275 "test.foo"(%i) : (index) -> ()276 }277 return278}279 280// -----281 282// CHECK-DAG: [[$UB_MAP:#map[0-9]*]] = affine_map<()[s0] -> ((-s0 + 10) ceildiv 4)>283// CHECK-DAG: [[$IV_MAP:#map[0-9]*]] = affine_map<(d0)[s0] -> (d0 * 4 + s0)>284 285// CHECK-LABEL: func @upper_bound_by_dim286func.func @upper_bound_by_dim(%arg0: index, %arg1: index) {287 // CHECK: affine.for %[[ARG0:.*]] = 0 to [[$UB_MAP]]()[%arg{{.*}}] {288 affine.for %i = affine_map<(d0, d1) -> (d0)>(%arg0, %arg1) to 10 step 4 {289 // CHECK-NEXT: %[[IV:.*]] = affine.apply [[$IV_MAP]](%[[ARG0]])[%arg{{.*}}]290 // CHECK-NEXT: "test.foo"(%[[IV]]) : (index) -> ()291 "test.foo"(%i) : (index) -> ()292 }293 return294}295 296// -----297 298// CHECK: [[$MAP:#map[0-9]*]] = affine_map<(d0) -> (d0 * 64)>299// CHECK: [[$MAP1:#map[0-9]*]] = affine_map<(d0) -> (2, (-d0 + 1024) ceildiv 32)>300// CHECK: [[$MAP2:#map[0-9]*]] = affine_map<(d0, d1) -> (d0 + d1 * 32)>301// CHECK: [[$MAP3:#map[0-9]*]] = affine_map<(d0, d1) -> (32, d0 - d1 + 64, -d1 + 1024)>302// CHECK: [[$MAP4:#map[0-9]*]] = affine_map<(d0, d1) -> (d0 + d1)>303#map0 = affine_map<(d0) -> (d0)>304#map1 = affine_map<(d0) -> (d0 + 64, 1024)>305#map2 = affine_map<(d0, d1) -> (d1 + 32, d0 + 64, 1024)>306// CHECK-LABEL: @multi_level_tiled_matmul()307func.func @multi_level_tiled_matmul() {308 // CHECK-NEXT: %[[BUF:.*]] = memref.alloc() : memref<1024xf16>309 %0 = memref.alloc() : memref<1024xf16>310 affine.for %arg0 = 0 to 1024 step 64 {311 // CHECK-NEXT: affine.for %[[ARG0:.*]] = 0 to 16 {312 // CHECK-NEXT: %[[IV0:.*]] = affine.apply [[$MAP]](%[[ARG0]])313 affine.for %arg3 = #map0(%arg0) to min #map1(%arg0) step 32 {314 // CHECK-NEXT: affine.for %[[ARG1:.*]] = 0 to min [[$MAP1]](%[[IV0]]) {315 // CHECK-NEXT: %[[IV1:.*]] = affine.apply [[$MAP2]](%[[IV0]], %[[ARG1]])316 affine.for %arg6 = #map0(%arg3) to min #map2(%arg0, %arg3) {317 // CHECK-NEXT: affine.for %[[ARG2:.*]] = 0 to min [[$MAP3]](%[[IV0]], %[[IV1]]) {318 // CHECK-NEXT: %[[IV2:.*]] = affine.apply [[$MAP4]](%[[IV1]], %[[ARG2]])319 // CHECK-NEXT: affine.load %[[BUF]][%[[IV2]]] : memref<1024xf16>320 affine.load %0[%arg6] : memref<1024xf16>321 }322 }323 }324 return325}326