91 lines · plain
1// RUN: mlir-opt %s --linalg-fuse-elementwise-ops --sparse-reinterpret-map --sparsification | FileCheck %s2 3#SV = #sparse_tensor.encoding<{ map = (d0) -> (d0 : compressed) }>4 5#trait = {6 indexing_maps = [7 affine_map<(i) -> (i)>, // A8 affine_map<(i) -> (i)> // B (out)9 ],10 iterator_types = ["parallel"],11 doc = "B(i) = OP A(i)"12}13 14 15// CHECK-LABEL: func.func @sparse_fusion(16// CHECK-SAME: %[[VAL_0:.*]]: tensor<100xf64, #sparse>) -> tensor<100xf64> {17// CHECK-DAG: %[[VAL_1:.*]] = arith.constant true18// CHECK-DAG: %[[VAL_2:.*]] = arith.constant 1 : index19// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 0 : index20// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 0.000000e+00 : f6421// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 100 : index22// CHECK-DAG: %[[VAL_6:.*]] = arith.constant 1.000000e+00 : f6423// CHECK-DAG: %[[VAL_7:.*]] = arith.constant 1.000000e+02 : f6424// CHECK-DAG: %[[VAL_8:.*]] = tensor.empty() : tensor<100xf64>25// CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.positions %[[VAL_0]] {level = 0 : index} : tensor<100xf64, #sparse> to memref<?xindex>26// CHECK-DAG: %[[VAL_10:.*]] = sparse_tensor.coordinates %[[VAL_0]] {level = 0 : index} : tensor<100xf64, #sparse> to memref<?xindex>27// CHECK-DAG: %[[VAL_11:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<100xf64, #sparse> to memref<?xf64>28// CHECK-DAG: %[[VAL_12:.*]] = bufferization.to_buffer %[[VAL_8]] :29// CHECK-DAG: linalg.fill ins(%[[VAL_4]] : f64) outs(%[[VAL_12]] : memref<100xf64>)30// CHECK-DAG: %[[VAL_13:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_3]]] : memref<?xindex>31// CHECK-DAG: %[[VAL_14:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_2]]] : memref<?xindex>32// CHECK: %[[VAL_15:.*]]:2 = scf.while (%[[VAL_16:.*]] = %[[VAL_13]], %[[VAL_17:.*]] = %[[VAL_3]]) : (index, index) -> (index, index) {33// CHECK: %[[VAL_18:.*]] = arith.cmpi ult, %[[VAL_16]], %[[VAL_14]] : index34// CHECK: scf.condition(%[[VAL_18]]) %[[VAL_16]], %[[VAL_17]] : index, index35// CHECK: } do {36// CHECK: ^bb0(%[[VAL_19:.*]]: index, %[[VAL_20:.*]]: index):37// CHECK: %[[VAL_21:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_19]]] : memref<?xindex>38// CHECK: %[[VAL_22:.*]] = arith.cmpi eq, %[[VAL_21]], %[[VAL_20]] : index39// CHECK: scf.if %[[VAL_22]] {40// CHECK: %[[VAL_23:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_19]]] : memref<?xf64>41// CHECK: %[[VAL_24:.*]] = arith.addf %[[VAL_23]], %[[VAL_6]] : f6442// CHECK: %[[VAL_25:.*]] = math.exp %[[VAL_24]] : f6443// CHECK: %[[VAL_26:.*]] = arith.maximumf %[[VAL_25]], %[[VAL_7]] : f6444// CHECK: memref.store %[[VAL_26]], %[[VAL_12]]{{\[}}%[[VAL_20]]] : memref<100xf64>45// CHECK: } else {46// CHECK: scf.if %[[VAL_1]] {47// CHECK: memref.store %[[VAL_7]], %[[VAL_12]]{{\[}}%[[VAL_20]]] : memref<100xf64>48// CHECK: } else {49// CHECK: }50// CHECK: }51// CHECK: %[[VAL_27:.*]] = arith.cmpi eq, %[[VAL_21]], %[[VAL_20]] : index52// CHECK: %[[VAL_28:.*]] = arith.addi %[[VAL_19]], %[[VAL_2]] : index53// CHECK: %[[VAL_29:.*]] = arith.select %[[VAL_27]], %[[VAL_28]], %[[VAL_19]] : index54// CHECK: %[[VAL_30:.*]] = arith.addi %[[VAL_20]], %[[VAL_2]] : index55// CHECK: scf.yield %[[VAL_29]], %[[VAL_30]] : index, index56// CHECK: }57// CHECK: scf.for %[[VAL_31:.*]] = %[[VAL_32:.*]]#1 to %[[VAL_5]] step %[[VAL_2]] {58// CHECK: memref.store %[[VAL_7]], %[[VAL_12]]{{\[}}%[[VAL_31]]] : memref<100xf64>59// CHECK: }60// CHECK: %[[VAL_33:.*]] = bufferization.to_tensor %[[VAL_12]] :61// CHECK: return %[[VAL_33]] : tensor<100xf64>62// CHECK: }63func.func @sparse_fusion(%argA: tensor<100xf64, #SV>) -> tensor<100xf64> {64 %c1 = arith.constant 1.0 : f6465 %c100 = arith.constant 100.0 : f6466 67 %t0 = tensor.empty() : tensor<100xf64>68 %l0 = linalg.generic #trait69 ins(%argA: tensor<100xf64, #SV>) outs(%t0: tensor<100xf64>) {70 ^bb0(%in0: f64, %out0: f64):71 %b0 = arith.addf %in0, %c1 : f6472 linalg.yield %b0 : f6473 } -> tensor<100xf64>74 %t1 = tensor.empty() : tensor<100xf64>75 %l1 = linalg.generic #trait76 ins(%l0: tensor<100xf64>) outs(%t1: tensor<100xf64>) {77 ^bb0(%in1: f64, %out1: f64):78 %b1 = math.exp %in1 : f6479 linalg.yield %b1 : f6480 } -> tensor<100xf64>81 %t2 = tensor.empty() : tensor<100xf64>82 %l2 = linalg.generic #trait83 ins(%l1: tensor<100xf64>) outs(%t2: tensor<100xf64>) {84 ^bb0(%in2: f64, %out2: f64):85 %b2 = arith.maximumf %in2, %c100 : f6486 linalg.yield %b2 : f6487 } -> tensor<100xf64>88 89 return %l2 : tensor<100xf64>90}91