63 lines · plain
1// RUN: mlir-opt %s --sparse-reinterpret-map -sparsification="parallelization-strategy=any-storage-any-loop" | \2// RUN: FileCheck %s3 4#CSR = #sparse_tensor.encoding<{5 map = (d0, d1) -> (d0 : dense, d1 : compressed)6}>7 8#trait_matvec = {9 indexing_maps = [10 affine_map<(i,j) -> (i,j)>, // A11 affine_map<(i,j) -> (j)>, // b12 affine_map<(i,j) -> (i)> // x (out)13 ],14 iterator_types = ["parallel", "reduction"],15 doc = "x(i) += A(i,j) * b(j)"16}17// CHECK-LABEL: func.func @matvec(18// CHECK-SAME: %[[TMP_arg0:.*]]: tensor<16x32xf32, #sparse{{[0-9]*}}>,19// CHECK-SAME: %[[TMP_arg1:.*]]: tensor<32xf32>,20// CHECK-SAME: %[[TMP_arg2:.*]]: tensor<16xf32>) -> tensor<16xf32> {21// CHECK-DAG: %[[TMP_c16:.*]] = arith.constant 16 : index22// CHECK-DAG: %[[TMP_c0:.*]] = arith.constant 0 : index23// CHECK-DAG: %[[TMP_c1:.*]] = arith.constant 1 : index24// CHECK-DAG: %[[TMP_0:.*]] = sparse_tensor.positions %[[TMP_arg0]] {level = 1 : index}25// CHECK-DAG: %[[TMP_1:.*]] = sparse_tensor.coordinates %[[TMP_arg0]] {level = 1 : index}26// CHECK-DAG: %[[TMP_2:.*]] = sparse_tensor.values %[[TMP_arg0]]27// CHECK-DAG: %[[TMP_3:.*]] = bufferization.to_buffer %[[TMP_arg1]] : tensor<32xf32> to memref<32xf32>28// CHECK-DAG: %[[TMP_4:.*]] = bufferization.to_buffer %[[TMP_arg2]] : tensor<16xf32> to memref<16xf32>29// CHECK: scf.parallel (%[[TMP_arg3:.*]]) = (%[[TMP_c0]]) to (%[[TMP_c16]]) step (%[[TMP_c1]]) {30// CHECK: %[[TMP_6:.*]] = memref.load %[[TMP_4]][%[[TMP_arg3]]] : memref<16xf32>31// CHECK: %[[TMP_7:.*]] = memref.load %[[TMP_0]][%[[TMP_arg3]]] : memref<?xindex>32// CHECK: %[[TMP_8:.*]] = arith.addi %[[TMP_arg3]], %[[TMP_c1]] : index33// CHECK: %[[TMP_9:.*]] = memref.load %[[TMP_0]][%[[TMP_8]]] : memref<?xindex>34// CHECK: %[[TMP_10:.*]] = scf.parallel (%[[TMP_arg4:.*]]) = (%[[TMP_7]]) to (%[[TMP_9]]) step (%[[TMP_c1]]) init (%[[TMP_6]]) -> f32 {35// CHECK: %[[TMP_11:.*]] = memref.load %[[TMP_1]][%[[TMP_arg4]]] : memref<?xindex>36// CHECK: %[[TMP_12:.*]] = memref.load %[[TMP_2]][%[[TMP_arg4]]] : memref<?xf32>37// CHECK: %[[TMP_13:.*]] = memref.load %[[TMP_3]][%[[TMP_11]]] : memref<32xf32>38// CHECK: %[[TMP_14:.*]] = arith.mulf %[[TMP_12]], %[[TMP_13]] : f3239// CHECK: scf.reduce(%[[TMP_14]] : f32) {40// CHECK: ^bb0(%[[TMP_arg5:.*]]: f32, %[[TMP_arg6:.*]]: f32):41// CHECK: %[[TMP_15:.*]] = arith.addf %[[TMP_arg5]], %[[TMP_arg6]] : f3242// CHECK: scf.reduce.return %[[TMP_15]] : f3243// CHECK: }44// CHECK: }45// CHECK: memref.store %[[TMP_10]], %[[TMP_4]][%[[TMP_arg3]]] : memref<16xf32>46// CHECK: scf.reduce47// CHECK: }48// CHECK: %[[TMP_5:.*]] = bufferization.to_tensor %[[TMP_4]] : memref<16xf32>49// CHECK: return %[[TMP_5]] : tensor<16xf32>50func.func @matvec(%arga: tensor<16x32xf32, #CSR>,51 %argb: tensor<32xf32>,52 %argx: tensor<16xf32>) -> tensor<16xf32> {53 %0 = linalg.generic #trait_matvec54 ins(%arga, %argb : tensor<16x32xf32, #CSR>, tensor<32xf32>)55 outs(%argx: tensor<16xf32>) {56 ^bb(%A: f32, %b: f32, %x: f32):57 %0 = arith.mulf %A, %b : f3258 %1 = arith.addf %0, %x : f3259 linalg.yield %1 : f3260 } -> tensor<16xf32>61 return %0 : tensor<16xf32>62}63