253 lines · plain
1// RUN: mlir-opt %s --transform-interpreter --split-input-file | FileCheck %s2 3// CHECK-DAG: #[[$map_p4:.*]] = affine_map<()[s0] -> (s0 + 4)>4// CHECK-DAG: #[[$map_p8:.*]] = affine_map<()[s0] -> (s0 + 8)>5// CHECK-DAG: #[[$bounds_map_4:.*]] = affine_map<(d0, d1, d2) -> (d0 - d1, 4)>6// CHECK-DAG: #[[$bounds_map_8:.*]] = affine_map<(d0, d1, d2) -> (d0 - d1, 8)>7 8// CHECK-LABEL: split_vector_transfer_read_2d(9// CHECK-SAME: %[[A:[a-zA-Z0-9_]*]]: memref10// CHECK-SAME: %[[i:[a-zA-Z0-9_]*]]: index11// CHECK-SAME: %[[j:[a-zA-Z0-9_]*]]: index12func.func @split_vector_transfer_read_2d(%A: memref<?x8xf32>, %i: index, %j: index) -> vector<4x8xf32> {13 %c0 = arith.constant 0 : index14 %f0 = arith.constant 0.0 : f3215 16 // CHECK-DAG: %[[c0:.*]] = arith.constant 0 : index17 // CHECK-DAG: %[[c4:.*]] = arith.constant 4 : index18 // CHECK-DAG: %[[c8:.*]] = arith.constant 8 : index19 // alloca for boundary full tile20 // CHECK: %[[alloc:.*]] = memref.alloca() {alignment = 32 : i64} : memref<4x8xf32>21 // %i + 4 <= dim(%A, 0)22 // CHECK: %[[idx0:.*]] = affine.apply #[[$map_p4]]()[%[[i]]]23 // CHECK: %[[d0:.*]] = memref.dim %[[A]], %[[c0]] : memref<?x8xf32>24 // CHECK: %[[cmp0:.*]] = arith.cmpi sle, %[[idx0]], %[[d0]] : index25 // %j + 8 <= dim(%A, 1)26 // CHECK: %[[idx1:.*]] = affine.apply #[[$map_p8]]()[%[[j]]]27 // CHECK: %[[cmp1:.*]] = arith.cmpi sle, %[[idx1]], %[[c8]] : index28 // are both conds true29 // CHECK: %[[cond:.*]] = arith.andi %[[cmp0]], %[[cmp1]] : i130 // CHECK: %[[ifres:.*]]:3 = scf.if %[[cond]] -> (memref<?x8xf32>, index, index) {31 // inBounds, just yield %A32 // CHECK: scf.yield %[[A]], %[[i]], %[[j]] : memref<?x8xf32>, index, index33 // CHECK: } else {34 // slow path, fill tmp alloc and yield a memref_casted version of it35 // CHECK: linalg.fill ins(%cst : f32) outs(%[[alloc]] : memref<4x8xf32>)36 // CHECK: %[[d0:.*]] = memref.dim %[[A]], %[[c0]] : memref<?x8xf32>37 // CHECK: %[[sv0:.*]] = affine.min #[[$bounds_map_4]](%[[d0]], %[[i]], %[[c4]])38 // CHECK: %[[sv1:.*]] = affine.min #[[$bounds_map_8]](%[[c8]], %[[j]], %[[c8]])39 // CHECK: %[[sv:.*]] = memref.subview %[[A]][%[[i]], %[[j]]] [%[[sv0]], %[[sv1]]] [1, 1]40 // CHECK-SAME: memref<?x8xf32> to memref<?x?xf32, strided<[8, 1], offset: ?>>41 // CHECK: %[[alloc_view:.*]] = memref.subview %[[alloc]][0, 0] [%[[sv0]], %[[sv1]]] [1, 1]42 // CHECK: memref.copy %[[sv]], %[[alloc_view]] : memref<?x?xf32, strided<[8, 1], offset: ?>> to memref<?x?xf32, strided{{.*}}>43 // CHECK: %[[yielded:.*]] = memref.cast %[[alloc]] :44 // CHECK-SAME: memref<4x8xf32> to memref<?x8xf32>45 // CHECK: scf.yield %[[yielded]], %[[c0]], %[[c0]] :46 // CHECK-SAME: memref<?x8xf32>, index, index47 // CHECK: }48 // CHECK: %[[res:.*]] = vector.transfer_read %[[ifres]]#0[%[[ifres]]#1, %[[ifres]]#2], %cst49 // CHECK-SAME: {in_bounds = [true, true]} : memref<?x8xf32>, vector<4x8xf32>50 %1 = vector.transfer_read %A[%i, %j], %f0 : memref<?x8xf32>, vector<4x8xf32>51 52 // CHECK: return %[[res]] : vector<4x8xf32>53 return %1: vector<4x8xf32>54}55 56// CHECK-LABEL: split_vector_transfer_read_strided_2d(57// CHECK-SAME: %[[A:[a-zA-Z0-9_]*]]: memref58// CHECK-SAME: %[[i:[a-zA-Z0-9_]*]]: index59// CHECK-SAME: %[[j:[a-zA-Z0-9_]*]]: index60func.func @split_vector_transfer_read_strided_2d(61 %A: memref<7x8xf32, strided<[?, 1], offset: ?>>,62 %i: index, %j: index) -> vector<4x8xf32> {63 %c0 = arith.constant 0 : index64 %f0 = arith.constant 0.0 : f3265 66 67 // CHECK-DAG: %[[c0:.*]] = arith.constant 0 : index68 // CHECK-DAG: %[[c4:.*]] = arith.constant 4 : index69 // CHECK-DAG: %[[c7:.*]] = arith.constant 7 : index70 // CHECK-DAG: %[[c8:.*]] = arith.constant 8 : index71 // alloca for boundary full tile72 // CHECK: %[[alloc:.*]] = memref.alloca() {alignment = 32 : i64} : memref<4x8xf32>73 // %i + 4 <= dim(%A, 0)74 // CHECK: %[[idx0:.*]] = affine.apply #[[$map_p4]]()[%[[i]]]75 // CHECK: %[[cmp0:.*]] = arith.cmpi sle, %[[idx0]], %[[c7]] : index76 // %j + 8 <= dim(%A, 1)77 // CHECK: %[[idx1:.*]] = affine.apply #[[$map_p8]]()[%[[j]]]78 // CHECK: %[[cmp1:.*]] = arith.cmpi sle, %[[idx1]], %[[c8]] : index79 // are both conds true80 // CHECK: %[[cond:.*]] = arith.andi %[[cmp0]], %[[cmp1]] : i181 // CHECK: %[[ifres:.*]]:3 = scf.if %[[cond]] -> (memref<?x8xf32, strided<[?, 1], offset: ?>>, index, index) {82 // inBounds but not cast-compatible: yield a memref_casted form of %A83 // CHECK: %[[casted:.*]] = memref.cast %arg0 :84 // CHECK-SAME: memref<7x8xf32, strided<[?, 1], offset: ?>> to memref<?x8xf32, strided<[?, 1], offset: ?>>85 // CHECK: scf.yield %[[casted]], %[[i]], %[[j]] :86 // CHECK-SAME: memref<?x8xf32, strided<[?, 1], offset: ?>>, index, index87 // CHECK: } else {88 // slow path, fill tmp alloc and yield a memref_casted version of it89 // CHECK: linalg.fill ins(%cst : f32) outs(%[[alloc]] : memref<4x8xf32>)90 // CHECK: %[[sv0:.*]] = affine.min #[[$bounds_map_4]](%[[c7]], %[[i]], %[[c4]])91 // CHECK: %[[sv1:.*]] = affine.min #[[$bounds_map_8]](%[[c8]], %[[j]], %[[c8]])92 // CHECK: %[[sv:.*]] = memref.subview %[[A]][%[[i]], %[[j]]] [%[[sv0]], %[[sv1]]] [1, 1]93 // CHECK-SAME: memref<7x8xf32, strided<[?, 1], offset: ?>> to memref<?x?xf32, strided<[?, 1], offset: ?>>94 // CHECK: %[[alloc_view:.*]] = memref.subview %[[alloc]][0, 0] [%[[sv0]], %[[sv1]]] [1, 1]95 // CHECK: memref.copy %[[sv]], %[[alloc_view]] : memref<?x?xf32, strided<[?, 1], offset: ?>> to memref<?x?xf32, strided{{.*}}>96 // CHECK: %[[yielded:.*]] = memref.cast %[[alloc]] :97 // CHECK-SAME: memref<4x8xf32> to memref<?x8xf32, strided<[?, 1], offset: ?>>98 // CHECK: scf.yield %[[yielded]], %[[c0]], %[[c0]] :99 // CHECK-SAME: memref<?x8xf32, strided<[?, 1], offset: ?>>, index, index100 // CHECK: }101 // CHECK: %[[res:.*]] = vector.transfer_read {{.*}} {in_bounds = [true, true]} :102 // CHECK-SAME: memref<?x8xf32, strided<[?, 1], offset: ?>>, vector<4x8xf32>103 %1 = vector.transfer_read %A[%i, %j], %f0 :104 memref<7x8xf32, strided<[?, 1], offset: ?>>, vector<4x8xf32>105 106 return %1 : vector<4x8xf32>107}108 109module attributes {transform.with_named_sequence} {110 transform.named_sequence @__transform_main(%root : !transform.any_op {transform.readonly}) {111 %func_op = transform.structured.match ops{["func.func"]} in %root : (!transform.any_op) -> !transform.op<"func.func">112 transform.apply_patterns to %func_op {113 transform.apply_patterns.vector.split_transfer_full_partial split_transfer_strategy = "linalg-copy"114 } : !transform.op<"func.func">115 transform.yield116 }117}118 119// -----120 121func.func @split_vector_transfer_write_2d(%V: vector<4x8xf32>, %A: memref<?x8xf32>, %i: index, %j: index) {122 vector.transfer_write %V, %A[%i, %j] :123 vector<4x8xf32>, memref<?x8xf32>124 return125}126 127// CHECK-DAG: #[[$MAP0:.*]] = affine_map<()[s0] -> (s0 + 4)>128// CHECK-DAG: #[[$MAP1:.*]] = affine_map<()[s0] -> (s0 + 8)>129// CHECK-DAG: #[[$MAP2:.*]] = affine_map<(d0, d1, d2) -> (d0 - d1, 4)>130// CHECK-DAG: #[[$MAP3:.*]] = affine_map<(d0, d1, d2) -> (d0 - d1, 8)>131 132// CHECK-LABEL: func @split_vector_transfer_write_2d(133// CHECK-SAME: %[[VEC:.*]]: vector<4x8xf32>,134// CHECK-SAME: %[[DEST:.*]]: memref<?x8xf32>,135// CHECK-SAME: %[[I:.*]]: index,136// CHECK-SAME: %[[J:.*]]: index) {137// CHECK-DAG: %[[CT:.*]] = arith.constant true138// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index139// CHECK-DAG: %[[C4:.*]] = arith.constant 4 : index140// CHECK-DAG: %[[C8:.*]] = arith.constant 8 : index141// CHECK: %[[TEMP:.*]] = memref.alloca() {alignment = 32 : i64} : memref<4x8xf32>142// CHECK: %[[IDX0:.*]] = affine.apply #[[$MAP0]]()[%[[I]]]143// CHECK: %[[DIM0:.*]] = memref.dim %[[DEST]], %[[C0]] : memref<?x8xf32>144// CHECK: %[[DIM0_IN:.*]] = arith.cmpi sle, %[[IDX0]], %[[DIM0]] : index145// CHECK: %[[DIM1:.*]] = affine.apply #[[$MAP1]]()[%[[J]]]146// CHECK: %[[DIM1_IN:.*]] = arith.cmpi sle, %[[DIM1]], %[[C8]] : index147// CHECK: %[[IN_BOUNDS:.*]] = arith.andi %[[DIM0_IN]], %[[DIM1_IN]] : i1148// CHECK: %[[IN_BOUND_DEST:.*]]:3 = scf.if %[[IN_BOUNDS]]149// CHECK-SAME: -> (memref<?x8xf32>, index, index) {150// CHECK: scf.yield %[[DEST]], %[[I]], %[[J]] : memref<?x8xf32>, index, index151// CHECK: } else {152// CHECK: %[[VAL_16:.*]] = memref.cast %[[TEMP]] : memref<4x8xf32> to memref<?x8xf32>153// CHECK: scf.yield %[[VAL_16]], %[[C0]], %[[C0]] : memref<?x8xf32>, index, index154// CHECK: }155// CHECK: vector.transfer_write %[[VEC]],156// CHECK-SAME: %[[IN_BOUND_DEST:.*]]#0[%[[IN_BOUND_DEST]]#1, %[[IN_BOUND_DEST]]#2]157// CHECK-SAME: {in_bounds = [true, true]} : vector<4x8xf32>, memref<?x8xf32>158// CHECK: %[[OUT_BOUNDS:.*]] = arith.xori %[[IN_BOUNDS]], %[[CT]] : i1159// CHECK: scf.if %[[OUT_BOUNDS]] {160// CHECK: %[[VAL_19:.*]] = memref.dim %[[DEST]], %[[C0]] : memref<?x8xf32>161// CHECK-DAG: %[[VAL_20:.*]] = affine.min #[[$MAP2]](%[[VAL_19]], %[[I]], %[[C4]])162// CHECK-DAG: %[[VAL_21:.*]] = affine.min #[[$MAP3]](%[[C8]], %[[J]], %[[C8]])163// CHECK: %[[VAL_22:.*]] = memref.subview %[[TEMP]]164// CHECK-SAME: [%[[I]], %[[J]]] [%[[VAL_20]], %[[VAL_21]]]165// CHECK-SAME: [1, 1] : memref<4x8xf32> to memref<?x?xf32, strided<[8, 1], offset: ?>>166// CHECK: %[[DEST_VIEW:.*]] = memref.subview %[[DEST]][0, 0] [%[[VAL_20]], %[[VAL_21]]] [1, 1]167// CHECK: memref.copy %[[VAL_22]], %[[DEST_VIEW]]168// CHECK-SAME: : memref<?x?xf32, strided<[8, 1], offset: ?>> to memref<?x?xf32, strided{{.*}}>169// CHECK: }170// CHECK: return171// CHECK: }172 173module attributes {transform.with_named_sequence} {174 transform.named_sequence @__transform_main(%root : !transform.any_op {transform.readonly}) {175 %func_op = transform.structured.match ops{["func.func"]} in %root : (!transform.any_op) -> !transform.op<"func.func">176 transform.apply_patterns to %func_op {177 transform.apply_patterns.vector.split_transfer_full_partial split_transfer_strategy = "linalg-copy"178 } : !transform.op<"func.func">179 transform.yield180 }181}182 183// -----184 185func.func @split_vector_transfer_write_strided_2d(186 %V: vector<4x8xf32>, %A: memref<7x8xf32, strided<[?, 1], offset: ?>>,187 %i: index, %j: index) {188 vector.transfer_write %V, %A[%i, %j] :189 vector<4x8xf32>, memref<7x8xf32, strided<[?, 1], offset: ?>>190 return191}192 193// CHECK-DAG: #[[$MAP1:.*]] = affine_map<()[s0] -> (s0 + 4)>194// CHECK-DAG: #[[$MAP2:.*]] = affine_map<()[s0] -> (s0 + 8)>195// CHECK-DAG: #[[$MAP3:.*]] = affine_map<(d0, d1, d2) -> (d0 - d1, 4)>196// CHECK-DAG: #[[$MAP4:.*]] = affine_map<(d0, d1, d2) -> (d0 - d1, 8)>197// CHECK-LABEL: func @split_vector_transfer_write_strided_2d(198// CHECK-SAME: %[[VEC:.*]]: vector<4x8xf32>,199// CHECK-SAME: %[[DEST:.*]]: memref<7x8xf32, strided<[?, 1], offset: ?>>,200// CHECK-SAME: %[[I:.*]]: index,201// CHECK-SAME: %[[J:.*]]: index) {202// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index203// CHECK-DAG: %[[CT:.*]] = arith.constant true204// CHECK-DAG: %[[C7:.*]] = arith.constant 7 : index205// CHECK-DAG: %[[C4:.*]] = arith.constant 4 : index206// CHECK-DAG: %[[C8:.*]] = arith.constant 8 : index207// CHECK: %[[TEMP:.*]] = memref.alloca() {alignment = 32 : i64} : memref<4x8xf32>208// CHECK: %[[DIM0:.*]] = affine.apply #[[$MAP1]]()[%[[I]]]209// CHECK: %[[DIM0_IN:.*]] = arith.cmpi sle, %[[DIM0]], %[[C7]] : index210// CHECK: %[[DIM1:.*]] = affine.apply #[[$MAP2]]()[%[[J]]]211// CHECK: %[[DIM1_IN:.*]] = arith.cmpi sle, %[[DIM1]], %[[C8]] : index212// CHECK: %[[IN_BOUNDS:.*]] = arith.andi %[[DIM0_IN]], %[[DIM1_IN]] : i1213// CHECK: %[[IN_BOUND_DEST:.*]]:3 = scf.if %[[IN_BOUNDS]]214// CHECK-SAME: -> (memref<?x8xf32, strided<[?, 1], offset: ?>>, index, index) {215// CHECK: %[[VAL_16:.*]] = memref.cast %[[DEST]]216// CHECK-SAME: : memref<7x8xf32, strided<[?, 1], offset: ?>> to memref<?x8xf32, strided<[?, 1], offset: ?>>217// CHECK: scf.yield %[[VAL_16]], %[[I]], %[[J]]218// CHECK-SAME: : memref<?x8xf32, strided<[?, 1], offset: ?>>, index, index219// CHECK: } else {220// CHECK: %[[VAL_17:.*]] = memref.cast %[[TEMP]]221// CHECK-SAME: : memref<4x8xf32> to memref<?x8xf32, strided<[?, 1], offset: ?>>222// CHECK: scf.yield %[[VAL_17]], %[[C0]], %[[C0]]223// CHECK-SAME: : memref<?x8xf32, strided<[?, 1], offset: ?>>, index, index224// CHECK: }225// CHECK: vector.transfer_write %[[VEC]],226// CHECK-SAME: %[[IN_BOUND_DEST:.*]]#0227// CHECK-SAME: [%[[IN_BOUND_DEST]]#1, %[[IN_BOUND_DEST]]#2]228// CHECK-SAME: {in_bounds = [true, true]}229// CHECK-SAME: : vector<4x8xf32>, memref<?x8xf32, strided<[?, 1], offset: ?>>230// CHECK: %[[OUT_BOUNDS:.*]] = arith.xori %[[IN_BOUNDS]], %[[CT]] : i1231// CHECK: scf.if %[[OUT_BOUNDS]] {232// CHECK-DAG: %[[VAL_20:.*]] = affine.min #[[$MAP3]](%[[C7]], %[[I]], %[[C4]])233// CHECK-DAG: %[[VAL_21:.*]] = affine.min #[[$MAP4]](%[[C8]], %[[J]], %[[C8]])234// CHECK: %[[VAL_22:.*]] = memref.subview %[[TEMP]]235// CHECK-SAME: [%[[I]], %[[J]]] [%[[VAL_20]], %[[VAL_21]]]236// CHECK-SAME: [1, 1] : memref<4x8xf32> to memref<?x?xf32, strided<[8, 1], offset: ?>>237// CHECK: %[[DEST_VIEW:.*]] = memref.subview %[[DEST]][0, 0] [%[[VAL_20]], %[[VAL_21]]] [1, 1]238// CHECK: memref.copy %[[VAL_22]], %[[DEST_VIEW]]239// CHECK-SAME: : memref<?x?xf32, strided<[8, 1], offset: ?>> to memref<?x?xf32, strided<[?, 1], offset: ?>>240// CHECK: }241// CHECK: return242// CHECK: }243 244module attributes {transform.with_named_sequence} {245 transform.named_sequence @__transform_main(%root : !transform.any_op {transform.readonly}) {246 %func_op = transform.structured.match ops{["func.func"]} in %root : (!transform.any_op) -> !transform.op<"func.func">247 transform.apply_patterns to %func_op {248 transform.apply_patterns.vector.split_transfer_full_partial split_transfer_strategy = "linalg-copy"249 } : !transform.op<"func.func">250 transform.yield251 }252}253