74 lines · plain
1// RUN: mlir-opt %s --sparse-reinterpret-map -sparsification | FileCheck %s2 3//4// A SDDMM implementation with "spy" function and5// in-place update of the sampling sparse matrix.6//7 8#SM = #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 : dense, d1 : compressed) }>9 10#trait_sampled_dense_dense = {11 indexing_maps = [12 affine_map<(i,j,k) -> (i,k)>, // A13 affine_map<(i,j,k) -> (k,j)>, // B14 affine_map<(i,j,k) -> (i,j)> // S15 ],16 iterator_types = ["parallel", "parallel", "reduction"],17 doc = "S(i,j) += spy[S(i,j)] x SUM_k A(i,k) B(k,j)"18}19 20// CHECK-LABEL: func.func @sparse_sampled_dd(21// CHECK-SAME: %[[VAL_0:.*0]]: tensor<8x8xf64>,22// CHECK-SAME: %[[VAL_1:.*1]]: tensor<8x8xf64>,23// CHECK-SAME: %[[VAL_2:.*2]]: tensor<8x8xf64, #sparse{{[0-9]*}}>) -> tensor<8x8xf64, #sparse{{[0-9]*}}> {24// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 8 : index25// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 0 : index26// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 1 : index27// CHECK-DAG: %[[VAL_6:.*]] = bufferization.to_buffer %[[VAL_0]] : tensor<8x8xf64> to memref<8x8xf64>28// CHECK-DAG: %[[VAL_7:.*]] = bufferization.to_buffer %[[VAL_1]] : tensor<8x8xf64> to memref<8x8xf64>29// CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.positions %[[VAL_2]] {level = 1 : index} : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xindex>30// CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.coordinates %[[VAL_2]] {level = 1 : index} : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xindex>31// CHECK-DAG: %[[VAL_10:.*]] = sparse_tensor.values %[[VAL_2]] : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xf64>32// CHECK: scf.for %[[VAL_11:.*]] = %[[VAL_4]] to %[[VAL_3]] step %[[VAL_5]] {33// CHECK: scf.for %[[VAL_12:.*]] = %[[VAL_4]] to %[[VAL_3]] step %[[VAL_5]] {34// CHECK: %[[VAL_13:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_11]]] : memref<?xindex>35// CHECK: %[[VAL_14:.*]] = arith.addi %[[VAL_11]], %[[VAL_5]] : index36// CHECK: %[[VAL_15:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_14]]] : memref<?xindex>37// CHECK: scf.for %[[VAL_16:.*]] = %[[VAL_13]] to %[[VAL_15]] step %[[VAL_5]] {38// CHECK: %[[VAL_17:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_16]]] : memref<?xindex>39// CHECK: %[[VAL_18:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_16]]] : memref<?xf64>40// CHECK: %[[VAL_19:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_11]], %[[VAL_12]]] : memref<8x8xf64>41// CHECK: %[[VAL_20:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_12]], %[[VAL_17]]] : memref<8x8xf64>42// CHECK: %[[VAL_21:.*]] = arith.mulf %[[VAL_19]], %[[VAL_20]] : f6443// CHECK: %[[VAL_22:.*]] = arith.addf %[[VAL_18]], %[[VAL_21]] : f6444// CHECK: memref.store %[[VAL_22]], %[[VAL_10]]{{\[}}%[[VAL_16]]] : memref<?xf64>45// CHECK: } {"Emitted from" = "linalg.generic"}46// CHECK: } {"Emitted from" = "linalg.generic"}47// CHECK: } {"Emitted from" = "linalg.generic"}48// CHECK: %[[VAL_23:.*]] = sparse_tensor.load %[[VAL_2]] : tensor<8x8xf64, #sparse{{[0-9]*}}>49// CHECK: return %[[VAL_23]] : tensor<8x8xf64, #sparse{{[0-9]*}}>50// CHECK: }51func.func @sparse_sampled_dd(%argA: tensor<8x8xf64>,52 %argB: tensor<8x8xf64>,53 %argS: tensor<8x8xf64, #SM>) -> tensor<8x8xf64, #SM> {54 %f0 = arith.constant 0.0 : f6455 %result = linalg.generic #trait_sampled_dense_dense56 ins(%argA, %argB: tensor<8x8xf64>, tensor<8x8xf64>) outs(%argS: tensor<8x8xf64, #SM>) {57 ^bb(%a: f64, %b: f64, %s: f64):58 %u = sparse_tensor.unary %s : f64 to f6459 present={60 ^bb0(%p: f64):61 %mul = arith.mulf %a, %b : f6462 sparse_tensor.yield %mul : f6463 }64 absent={}65 %r = sparse_tensor.reduce %s, %u, %f0 : f64 {66 ^bb0(%p: f64, %q: f64):67 %add = arith.addf %p, %q : f6468 sparse_tensor.yield %add : f6469 }70 linalg.yield %r : f6471 } -> tensor<8x8xf64, #SM>72 return %result : tensor<8x8xf64, #SM>73}74