125 lines · plain
1// RUN: mlir-opt %s --linalg-generalize-named-ops \2// RUN: --sparse-reinterpret-map --sparsification --sparse-tensor-codegen \3// RUN: --canonicalize --cse | FileCheck %s4 5#CSR = #sparse_tensor.encoding<{6 map = (d0, d1) -> (d0 : dense, d1 : compressed)7}>8 9//10// Computes C = A x B with all matrices sparse (SpMSpM) in CSR.11//12// CHECK-LABEL: func.func @matmul(13// CHECK-SAME: %[[VAL_0:.*0]]: memref<?xindex>,14// CHECK-SAME: %[[VAL_1:.*1]]: memref<?xindex>,15// CHECK-SAME: %[[VAL_2:.*2]]: memref<?xf64>,16// CHECK-SAME: %[[VAL_3:.*3]]: !sparse_tensor.storage_specifier17// CHECK-SAME: %[[VAL_4:.*4]]: memref<?xindex>,18// CHECK-SAME: %[[VAL_5:.*5]]: memref<?xindex>,19// CHECK-SAME: %[[VAL_6:.*6]]: memref<?xf64>,20// CHECK-SAME: %[[VAL_7:.*7]]: !sparse_tensor.storage_specifier21// CHECK-DAG: %[[VAL_8:.*]] = arith.constant 0.000000e+00 : f6422// CHECK-DAG: %[[VAL_9:.*]] = arith.constant true23// CHECK-DAG: %[[VAL_10:.*]] = arith.constant false24// CHECK-DAG: %[[VAL_11:.*]] = arith.constant 1 : index25// CHECK-DAG: %[[VAL_12:.*]] = arith.constant 0 : index26// CHECK-DAG: %[[VAL_13:.*]] = arith.constant 4 : index27// CHECK: %[[VAL_14:.*]] = memref.alloc() : memref<16xindex>28// CHECK: %[[VAL_15:.*]] = memref.cast %[[VAL_14]] : memref<16xindex> to memref<?xindex>29// CHECK: %[[VAL_16:.*]] = memref.alloc() : memref<16xindex>30// CHECK: %[[VAL_17:.*]] = memref.cast %[[VAL_16]] : memref<16xindex> to memref<?xindex>31// CHECK: %[[VAL_18:.*]] = memref.alloc() : memref<16xf64>32// CHECK: %[[VAL_19:.*]] = memref.cast %[[VAL_18]] : memref<16xf64> to memref<?xf64>33// CHECK: %[[VAL_20:.*]] = sparse_tensor.storage_specifier.init : !sparse_tensor.storage_specifier34// CHECK: %[[VAL_21:.*]] = sparse_tensor.storage_specifier.set %[[VAL_20]] lvl_sz at 0 with %[[VAL_13]] : !sparse_tensor.storage_specifier35// CHECK: %[[VAL_22:.*]] = sparse_tensor.storage_specifier.set %[[VAL_21]] lvl_sz at 1 with %[[VAL_13]] : !sparse_tensor.storage_specifier36// CHECK: %[[VAL_23:.*]] = sparse_tensor.storage_specifier.get %[[VAL_22]] pos_mem_sz at 1 : !sparse_tensor.storage_specifier37// CHECK: %[[VAL_24:.*]], %[[VAL_25:.*]] = sparse_tensor.push_back %[[VAL_23]], %[[VAL_15]], %[[VAL_12]] : index, memref<?xindex>, index38// CHECK: %[[VAL_26:.*]] = sparse_tensor.storage_specifier.set %[[VAL_22]] pos_mem_sz at 1 with %[[VAL_25]] : !sparse_tensor.storage_specifier39// CHECK: %[[VAL_27:.*]], %[[VAL_28:.*]] = sparse_tensor.push_back %[[VAL_25]], %[[VAL_24]], %[[VAL_12]], %[[VAL_13]] : index, memref<?xindex>, index, index40// CHECK: %[[VAL_29:.*]] = sparse_tensor.storage_specifier.set %[[VAL_26]] pos_mem_sz at 1 with %[[VAL_28]] : !sparse_tensor.storage_specifier41// CHECK: %[[VAL_30:.*]] = memref.alloc() : memref<4xf64>42// CHECK: %[[VAL_31:.*]] = memref.alloc() : memref<4xi1>43// CHECK: %[[VAL_32:.*]] = memref.alloc() : memref<4xindex>44// CHECK: %[[VAL_33:.*]] = memref.cast %[[VAL_32]] : memref<4xindex> to memref<?xindex>45// CHECK: linalg.fill ins(%[[VAL_8]] : f64) outs(%[[VAL_30]] : memref<4xf64>)46// CHECK: linalg.fill ins(%[[VAL_10]] : i1) outs(%[[VAL_31]] : memref<4xi1>)47// CHECK-DAG: %[[VAL_34:.*]] = sparse_tensor.storage_specifier.get %[[VAL_3]] pos_mem_sz at 1 : !sparse_tensor.storage_specifier48// CHECK-DAG: %[[VAL_35:.*]] = memref.subview %[[VAL_0]][0] {{\[}}%[[VAL_34]]] [1] : memref<?xindex> to memref<?xindex>49// CHECK-DAG: %[[VAL_36:.*]] = sparse_tensor.storage_specifier.get %[[VAL_3]] crd_mem_sz at 1 : !sparse_tensor.storage_specifier50// CHECK-DAG: %[[VAL_37:.*]] = memref.subview %[[VAL_1]][0] {{\[}}%[[VAL_36]]] [1] : memref<?xindex> to memref<?xindex>51// CHECK-DAG: %[[VAL_38:.*]] = sparse_tensor.storage_specifier.get %[[VAL_3]] val_mem_sz : !sparse_tensor.storage_specifier52// CHECK-DAG: %[[VAL_39:.*]] = memref.subview %[[VAL_2]][0] {{\[}}%[[VAL_38]]] [1] : memref<?xf64> to memref<?xf64>53// CHECK-DAG: %[[VAL_40:.*]] = sparse_tensor.storage_specifier.get %[[VAL_7]] pos_mem_sz at 1 : !sparse_tensor.storage_specifier54// CHECK-DAG: %[[VAL_41:.*]] = memref.subview %[[VAL_4]][0] {{\[}}%[[VAL_40]]] [1] : memref<?xindex> to memref<?xindex>55// CHECK-DAG: %[[VAL_42:.*]] = sparse_tensor.storage_specifier.get %[[VAL_7]] crd_mem_sz at 1 : !sparse_tensor.storage_specifier56// CHECK-DAG: %[[VAL_43:.*]] = memref.subview %[[VAL_5]][0] {{\[}}%[[VAL_42]]] [1] : memref<?xindex> to memref<?xindex>57// CHECK-DAG: %[[VAL_44:.*]] = sparse_tensor.storage_specifier.get %[[VAL_7]] val_mem_sz : !sparse_tensor.storage_specifier58// CHECK-DAG: %[[VAL_45:.*]] = memref.subview %[[VAL_6]][0] {{\[}}%[[VAL_44]]] [1] : memref<?xf64> to memref<?xf64>59// CHECK: %[[VAL_46:.*]]:4 = scf.for %[[VAL_47:.*]] = %[[VAL_12]] to %[[VAL_13]] step %[[VAL_11]] iter_args(%[[VAL_48:.*]] = %[[VAL_27]], %[[VAL_49:.*]] = %[[VAL_17]], %[[VAL_50:.*]] = %[[VAL_19]], %[[VAL_51:.*]] = %[[VAL_29]]) -> (memref<?xindex>, memref<?xindex>, memref<?xf64>, !sparse_tensor.storage_specifier60// CHECK: %[[VAL_52:.*]] = memref.load %[[VAL_35]]{{\[}}%[[VAL_47]]] : memref<?xindex>61// CHECK: %[[VAL_53:.*]] = arith.addi %[[VAL_47]], %[[VAL_11]] : index62// CHECK: %[[VAL_54:.*]] = memref.load %[[VAL_35]]{{\[}}%[[VAL_53]]] : memref<?xindex>63// CHECK: %[[VAL_55:.*]] = scf.for %[[VAL_56:.*]] = %[[VAL_52]] to %[[VAL_54]] step %[[VAL_11]] iter_args(%[[VAL_57:.*]] = %[[VAL_12]]) -> (index) {64// CHECK: %[[VAL_58:.*]] = memref.load %[[VAL_37]]{{\[}}%[[VAL_56]]] : memref<?xindex>65// CHECK: %[[VAL_59:.*]] = memref.load %[[VAL_39]]{{\[}}%[[VAL_56]]] : memref<?xf64>66// CHECK: %[[VAL_60:.*]] = memref.load %[[VAL_41]]{{\[}}%[[VAL_58]]] : memref<?xindex>67// CHECK: %[[VAL_61:.*]] = arith.addi %[[VAL_58]], %[[VAL_11]] : index68// CHECK: %[[VAL_62:.*]] = memref.load %[[VAL_41]]{{\[}}%[[VAL_61]]] : memref<?xindex>69// CHECK: %[[VAL_63:.*]] = scf.for %[[VAL_64:.*]] = %[[VAL_60]] to %[[VAL_62]] step %[[VAL_11]] iter_args(%[[VAL_65:.*]] = %[[VAL_57]]) -> (index) {70// CHECK: %[[VAL_66:.*]] = memref.load %[[VAL_43]]{{\[}}%[[VAL_64]]] : memref<?xindex>71// CHECK: %[[VAL_67:.*]] = memref.load %[[VAL_30]]{{\[}}%[[VAL_66]]] : memref<4xf64>72// CHECK: %[[VAL_68:.*]] = memref.load %[[VAL_45]]{{\[}}%[[VAL_64]]] : memref<?xf64>73// CHECK: %[[VAL_69:.*]] = arith.mulf %[[VAL_59]], %[[VAL_68]] : f6474// CHECK: %[[VAL_70:.*]] = arith.addf %[[VAL_67]], %[[VAL_69]] : f6475// CHECK: %[[VAL_71:.*]] = memref.load %[[VAL_31]]{{\[}}%[[VAL_66]]] : memref<4xi1>76// CHECK: %[[VAL_72:.*]] = arith.cmpi eq, %[[VAL_71]], %[[VAL_10]] : i177// CHECK: %[[VAL_73:.*]] = scf.if %[[VAL_72]] -> (index) {78// CHECK: memref.store %[[VAL_9]], %[[VAL_31]]{{\[}}%[[VAL_66]]] : memref<4xi1>79// CHECK: memref.store %[[VAL_66]], %[[VAL_32]]{{\[}}%[[VAL_65]]] : memref<4xindex>80// CHECK: %[[VAL_74:.*]] = arith.addi %[[VAL_65]], %[[VAL_11]] : index81// CHECK: scf.yield %[[VAL_74]] : index82// CHECK: } else {83// CHECK: scf.yield %[[VAL_65]] : index84// CHECK: }85// CHECK: memref.store %[[VAL_70]], %[[VAL_30]]{{\[}}%[[VAL_66]]] : memref<4xf64>86// CHECK: scf.yield %[[VAL_73]] : index87// CHECK: } {"Emitted from" = "linalg.generic"}88// CHECK: scf.yield %[[VAL_63]] : index89// CHECK: } {"Emitted from" = "linalg.generic"}90// CHECK: sparse_tensor.sort hybrid_quick_sort %[[VAL_55]], %[[VAL_33]]91// CHECK: %[[VAL_75:.*]]:4 = scf.for %[[VAL_76:.*]] = %[[VAL_12]] to %[[VAL_55]] step %[[VAL_11]] iter_args(%[[VAL_77:.*]] = %[[VAL_48]], %[[VAL_78:.*]] = %[[VAL_49]], %[[VAL_79:.*]] = %[[VAL_50]], %[[VAL_80:.*]] = %[[VAL_51]]) -> (memref<?xindex>, memref<?xindex>, memref<?xf64>, !sparse_tensor.storage_specifier92// CHECK: %[[VAL_81:.*]] = memref.load %[[VAL_32]]{{\[}}%[[VAL_76]]] : memref<4xindex>93// CHECK: %[[VAL_82:.*]] = memref.load %[[VAL_30]]{{\[}}%[[VAL_81]]] : memref<4xf64>94// CHECK: %[[VAL_83:.*]]:4 = func.call @_insert_dense_compressed_4_4_f64_0_0(%[[VAL_77]], %[[VAL_78]], %[[VAL_79]], %[[VAL_80]], %[[VAL_47]], %[[VAL_81]], %[[VAL_82]]) : (memref<?xindex>, memref<?xindex>, memref<?xf64>, !sparse_tensor.storage_specifier95// CHECK: memref.store %[[VAL_8]], %[[VAL_30]]{{\[}}%[[VAL_81]]] : memref<4xf64>96// CHECK: memref.store %[[VAL_10]], %[[VAL_31]]{{\[}}%[[VAL_81]]] : memref<4xi1>97// CHECK: scf.yield %[[VAL_83]]#0, %[[VAL_83]]#1, %[[VAL_83]]#2, %[[VAL_83]]#3 : memref<?xindex>, memref<?xindex>, memref<?xf64>, !sparse_tensor.storage_specifier98// CHECK: }99// CHECK: scf.yield %[[VAL_84:.*]]#0, %[[VAL_84]]#1, %[[VAL_84]]#2, %[[VAL_84]]#3 : memref<?xindex>, memref<?xindex>, memref<?xf64>, !sparse_tensor.storage_specifier100// CHECK: } {"Emitted from" = "linalg.generic"}101// CHECK: memref.dealloc %[[VAL_30]] : memref<4xf64>102// CHECK: memref.dealloc %[[VAL_31]] : memref<4xi1>103// CHECK: memref.dealloc %[[VAL_32]] : memref<4xindex>104// CHECK: %[[VAL_85:.*]] = sparse_tensor.storage_specifier.get %[[VAL_86:.*]]#3 pos_mem_sz at 1 : !sparse_tensor.storage_specifier105// CHECK: %[[VAL_87:.*]] = memref.load %[[VAL_86]]#0{{\[}}%[[VAL_12]]] : memref<?xindex>106// CHECK: %[[VAL_88:.*]] = scf.for %[[VAL_89:.*]] = %[[VAL_11]] to %[[VAL_85]] step %[[VAL_11]] iter_args(%[[VAL_90:.*]] = %[[VAL_87]]) -> (index) {107// CHECK: %[[VAL_91:.*]] = memref.load %[[VAL_86]]#0{{\[}}%[[VAL_89]]] : memref<?xindex>108// CHECK: %[[VAL_92:.*]] = arith.cmpi eq, %[[VAL_91]], %[[VAL_12]] : index109// CHECK: %[[VAL_93:.*]] = arith.select %[[VAL_92]], %[[VAL_90]], %[[VAL_91]] : index110// CHECK: scf.if %[[VAL_92]] {111// CHECK: memref.store %[[VAL_90]], %[[VAL_86]]#0{{\[}}%[[VAL_89]]] : memref<?xindex>112// CHECK: }113// CHECK: scf.yield %[[VAL_93]] : index114// CHECK: }115// CHECK: return %[[VAL_86]]#0, %[[VAL_86]]#1, %[[VAL_86]]#2, %[[VAL_86]]#3 : memref<?xindex>, memref<?xindex>, memref<?xf64>, !sparse_tensor.storage_specifier116// CHECK: }117func.func @matmul(%A: tensor<4x8xf64, #CSR>,118 %B: tensor<8x4xf64, #CSR>) -> tensor<4x4xf64, #CSR> {119 %C = tensor.empty() : tensor<4x4xf64, #CSR>120 %D = linalg.matmul121 ins(%A, %B: tensor<4x8xf64, #CSR>, tensor<8x4xf64, #CSR>)122 outs(%C: tensor<4x4xf64, #CSR>) -> tensor<4x4xf64, #CSR>123 return %D: tensor<4x4xf64, #CSR>124}125