103 lines · plain
1// RUN: mlir-opt %s --pre-sparsification-rewrite --sparse-reinterpret-map --sparsification --cse | FileCheck %s2 3#SM = #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 : compressed, d1 : compressed) }>4 5#trait_matmul = {6 indexing_maps = [7 affine_map<(d0, d1, d2) -> (d1, d0)>,8 affine_map<(d0, d1, d2) -> (d0, d2)>,9 affine_map<(d0, d1, d2) -> (d1, d2)>10 ],11 iterator_types = ["reduction", "parallel", "parallel"]12}13 14#trait_scale = {15 indexing_maps = [16 affine_map<(d0, d1) -> (d0, d1)>,17 affine_map<(d0, d1) -> (d0, d1)>,18 affine_map<(d0, d1) -> (d0, d1)>19 ],20 iterator_types = ["parallel", "parallel"]21}22 23// CHECK-LABEL: func.func @sparse_sampled_dd_unfused(24// CHECK-SAME: %[[VAL_0:.*]]: tensor<8x8xf64, #sparse{{[0-9]*}}>,25// CHECK-SAME: %[[VAL_1:.*]]: tensor<8x8xf64>,26// CHECK-SAME: %[[VAL_2:.*]]: tensor<8x8xf64>) -> tensor<8x8xf64, #sparse{{[0-9]*}}> {27// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 8 : index28// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 0 : index29// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 1 : index30// CHECK-DAG: %[[VAL_6:.*]] = arith.constant false31// CHECK-DAG: %[[VAL_7:.*]] = arith.constant true32// CHECK-DAG: %[[VAL_8:.*]] = tensor.empty() : tensor<8x8xf64, #sparse{{[0-9]*}}>33// CHECK-DAG: %[[VAL_9:.*]] = bufferization.to_buffer %[[VAL_1]] : tensor<8x8xf64> to memref<8x8xf64>34// CHECK-DAG: %[[VAL_10:.*]] = bufferization.to_buffer %[[VAL_2]] : tensor<8x8xf64> to memref<8x8xf64>35// CHECK-DAG: %[[VAL_11:.*]] = sparse_tensor.positions %[[VAL_0]] {level = 0 : index} : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xindex>36// CHECK-DAG: %[[VAL_12:.*]] = sparse_tensor.coordinates %[[VAL_0]] {level = 0 : index} : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xindex>37// CHECK-DAG: %[[VAL_13:.*]] = sparse_tensor.positions %[[VAL_0]] {level = 1 : index} : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xindex>38// CHECK-DAG: %[[VAL_14:.*]] = sparse_tensor.coordinates %[[VAL_0]] {level = 1 : index} : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xindex>39// CHECK-DAG: %[[VAL_15:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xf64>40// CHECK: %[[VAL_16:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_4]]] : memref<?xindex>41// CHECK: %[[VAL_17:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_5]]] : memref<?xindex>42// CHECK: %[[VAL_18:.*]] = scf.for %[[VAL_19:.*]] = %[[VAL_16]] to %[[VAL_17]] step %[[VAL_5]] iter_args(%[[VAL_20:.*]] = %[[VAL_8]]) -> (tensor<8x8xf64, #sparse{{[0-9]*}}>) {43// CHECK: %[[VAL_21:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_19]]] : memref<?xindex>44// CHECK: %[[VAL_22:.*]], %[[VAL_23:.*]], %[[VAL_24:.*]], %[[VAL_25:.*]] = sparse_tensor.expand %[[VAL_8]] : tensor<8x8xf64, #sparse{{[0-9]*}}> to memref<?xf64>, memref<?xi1>, memref<?xindex>45// CHECK: %[[VAL_26:.*]] = scf.for %[[VAL_27:.*]] = %[[VAL_4]] to %[[VAL_3]] step %[[VAL_5]] iter_args(%[[VAL_28:.*]] = %[[VAL_25]]) -> (index) {46// CHECK: %[[VAL_29:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_21]], %[[VAL_27]]] : memref<8x8xf64>47// CHECK: %[[VAL_30:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_19]]] : memref<?xindex>48// CHECK: %[[VAL_31:.*]] = arith.addi %[[VAL_19]], %[[VAL_5]] : index49// CHECK: %[[VAL_32:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_31]]] : memref<?xindex>50// CHECK: %[[VAL_33:.*]] = scf.for %[[VAL_34:.*]] = %[[VAL_30]] to %[[VAL_32]] step %[[VAL_5]] iter_args(%[[VAL_35:.*]] = %[[VAL_28]]) -> (index) {51// CHECK: %[[VAL_36:.*]] = memref.load %[[VAL_14]]{{\[}}%[[VAL_34]]] : memref<?xindex>52// CHECK: %[[VAL_37:.*]] = memref.load %[[VAL_22]]{{\[}}%[[VAL_36]]] : memref<?xf64>53// CHECK: %[[VAL_38:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_27]], %[[VAL_36]]] : memref<8x8xf64>54// CHECK: %[[VAL_39:.*]] = arith.mulf %[[VAL_29]], %[[VAL_38]] : f6455// CHECK: %[[VAL_40:.*]] = memref.load %[[VAL_15]]{{\[}}%[[VAL_34]]] : memref<?xf64>56// CHECK: %[[VAL_41:.*]] = arith.mulf %[[VAL_39]], %[[VAL_40]] : f6457// CHECK: %[[VAL_42:.*]] = arith.addf %[[VAL_37]], %[[VAL_41]] : f6458// CHECK: %[[VAL_43:.*]] = memref.load %[[VAL_23]]{{\[}}%[[VAL_36]]] : memref<?xi1>59// CHECK: %[[VAL_44:.*]] = arith.cmpi eq, %[[VAL_43]], %[[VAL_6]] : i160// CHECK: %[[VAL_45:.*]] = scf.if %[[VAL_44]] -> (index) {61// CHECK: memref.store %[[VAL_7]], %[[VAL_23]]{{\[}}%[[VAL_36]]] : memref<?xi1>62// CHECK: memref.store %[[VAL_36]], %[[VAL_24]]{{\[}}%[[VAL_35]]] : memref<?xindex>63// CHECK: %[[VAL_46:.*]] = arith.addi %[[VAL_35]], %[[VAL_5]] : index64// CHECK: scf.yield %[[VAL_46]] : index65// CHECK: } else {66// CHECK: scf.yield %[[VAL_35]] : index67// CHECK: }68// CHECK: memref.store %[[VAL_42]], %[[VAL_22]]{{\[}}%[[VAL_36]]] : memref<?xf64>69// CHECK: scf.yield %[[VAL_47:.*]] : index70// CHECK: } {"Emitted from" = "linalg.generic"}71// CHECK: scf.yield %[[VAL_48:.*]] : index72// CHECK: } {"Emitted from" = "linalg.generic"}73// CHECK: %[[VAL_49:.*]] = sparse_tensor.compress %[[VAL_22]], %[[VAL_23]], %[[VAL_24]], %[[VAL_50:.*]] into %[[VAL_20]]{{\[}}%[[VAL_21]]] : memref<?xf64>, memref<?xi1>, memref<?xindex>, tensor<8x8xf64, #sparse{{[0-9]*}}>74// CHECK: scf.yield %[[VAL_49]] : tensor<8x8xf64, #sparse{{[0-9]*}}>75// CHECK: } {"Emitted from" = "linalg.generic"}76// CHECK: %[[VAL_51:.*]] = sparse_tensor.load %[[VAL_52:.*]] hasInserts : tensor<8x8xf64, #sparse{{[0-9]*}}>77// CHECK: return %[[VAL_51]] : tensor<8x8xf64, #sparse{{[0-9]*}}>78// CHECK: }79func.func @sparse_sampled_dd_unfused(%args: tensor<8x8xf64, #SM>,80 %arga: tensor<8x8xf64>,81 %argb: tensor<8x8xf64>) -> tensor<8x8xf64, #SM> {82 // Perform dense-dense matrix matrix multiplication.83 %1 = arith.constant dense<0.0> : tensor<8x8xf64>84 %2 = linalg.generic #trait_matmul85 ins(%arga, %argb : tensor<8x8xf64>, tensor<8x8xf64>)86 outs(%1 : tensor<8x8xf64>) {87 ^bb0(%a: f64, %b: f64, %x: f64):88 %p = arith.mulf %a, %b : f6489 %q = arith.addf %x, %p : f6490 linalg.yield %q : f6491 } -> tensor<8x8xf64>92 // Sample the result with elements-wise multiplication with sparse matrix.93 %3 = tensor.empty() : tensor<8x8xf64, #SM>94 %4 = linalg.generic #trait_scale95 ins(%2, %args : tensor<8x8xf64>, tensor<8x8xf64, #SM>)96 outs(%3 : tensor<8x8xf64, #SM>) {97 ^bb0(%t: f64, %s: f64, %x: f64):98 %r = arith.mulf %t, %s : f6499 linalg.yield %r : f64100 } -> tensor<8x8xf64, #SM>101 return %4 : tensor<8x8xf64, #SM>102}103