1// RUN: mlir-opt %s --sparsification-and-bufferization | FileCheck %s 2 3#map = affine_map<(d0, d1, d2) -> (d0, d1, d2)> 4 5#sparse = #sparse_tensor.encoding<{ 6 map = (d0, d1, d2) -> (d0 : dense, d1 : dense, d2 : compressed) 7}> 8 9// 10// Make sure a simple ReLU passes the sparsifier 11// 12// CHECK-LABEL: func.func @relu 13// CHECK: scf.for 14// CHECK: scf.for 15// CHECK: scf.for 16// CHECK: arith.cmpf ugt 17// CHECK: arith.select 18// 19func.func @relu(%arg0: tensor<10x20x30xf64, #sparse>) -> tensor<10x20x30xf64, #sparse> { 20 %cst = arith.constant 0.000000e+00 : f64 21 %0 = tensor.empty() : tensor<10x20x30xf64> 22 %1 = linalg.generic { 23 indexing_maps = [#map, #map], 24 iterator_types = ["parallel", "parallel", "parallel"]} 25 ins(%arg0 : tensor<10x20x30xf64, #sparse>) 26 outs(%0 : tensor<10x20x30xf64>) { 27 ^bb0(%in: f64, %out: f64): 28 %2 = arith.cmpf ugt, %in, %cst : f64 29 %3 = arith.select %2, %in, %cst : f64 30 linalg.yield %3 : f64 31 } -> tensor<10x20x30xf64> 32 %cast = tensor.cast %1 : tensor<10x20x30xf64> to tensor<10x20x30xf64, #sparse> 33 return %cast : tensor<10x20x30xf64, #sparse> 34} 35