1// NOTE: Assertions have been autogenerated by utils/generate-test-checks.py 2// RUN: mlir-opt %s \ 3// RUN: --linalg-generalize-named-ops --linalg-fuse-elementwise-ops \ 4// RUN: --sparsification | FileCheck %s 5 6#SparseVector = #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }> 7 8#DCSR = #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }> 9 10// CHECK-LABEL: func @matmul1( 11// CHECK-SAME: %[[VAL_0:.*]]: tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>>, 12// CHECK-SAME: %[[VAL_1:.*]]: tensor<20x30xf32>, 13// CHECK-SAME: %[[VAL_2:.*]]: tensor<10x30xf32>) -> tensor<10x30xf32> { 14// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 0 : index 15// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 1 : index 16// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 30 : index 17// CHECK-DAG: %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_0]], %[[VAL_3]] : tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>> 18// CHECK-DAG: %[[VAL_7:.*]] = sparse_tensor.indices %[[VAL_0]], %[[VAL_3]] : tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>> 19// CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_0]], %[[VAL_4]] : tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>> 20// CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_0]], %[[VAL_4]] : tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>> 21// CHECK-DAG: %[[VAL_10:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>> 22// CHECK-DAG: %[[VAL_11:.*]] = bufferization.to_memref %[[VAL_1]] : memref<20x30xf32> 23// CHECK-DAG: %[[VAL_13:.*]] = bufferization.to_memref %[[VAL_2]] : memref<10x30xf32> 24// CHECK: %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_3]]] : memref<?xindex> 25// CHECK: %[[VAL_15:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex> 26// CHECK: scf.for %[[VAL_16:.*]] = %[[VAL_14]] to %[[VAL_15]] step %[[VAL_4]] { 27// CHECK: %[[VAL_17:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_16]]] : memref<?xindex> 28// CHECK: %[[VAL_18:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_16]]] : memref<?xindex> 29// CHECK: %[[VAL_19:.*]] = arith.addi %[[VAL_16]], %[[VAL_4]] : index 30// CHECK: %[[VAL_20:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_19]]] : memref<?xindex> 31// CHECK: scf.for %[[VAL_21:.*]] = %[[VAL_18]] to %[[VAL_20]] step %[[VAL_4]] { 32// CHECK: %[[VAL_22:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_21]]] : memref<?xindex> 33// CHECK: %[[VAL_23:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_21]]] : memref<?xf32> 34// CHECK: scf.for %[[VAL_24:.*]] = %[[VAL_3]] to %[[VAL_5]] step %[[VAL_4]] { 35// CHECK: %[[VAL_25:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_17]], %[[VAL_24]]] : memref<10x30xf32> 36// CHECK: %[[VAL_26:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_22]], %[[VAL_24]]] : memref<20x30xf32> 37// CHECK: %[[VAL_27:.*]] = arith.mulf %[[VAL_23]], %[[VAL_26]] : f32 38// CHECK: %[[VAL_28:.*]] = arith.addf %[[VAL_25]], %[[VAL_27]] : f32 39// CHECK: memref.store %[[VAL_28]], %[[VAL_13]]{{\[}}%[[VAL_17]], %[[VAL_24]]] : memref<10x30xf32> 40// CHECK: } 41// CHECK: } 42// CHECK: } 43// CHECK: %[[VAL_29:.*]] = bufferization.to_tensor %[[VAL_13]] : memref<10x30xf32> 44// CHECK: return %[[VAL_29]] : tensor<10x30xf32> 45// CHECK: } 46func.func @matmul1(%a: tensor<10x20xf32, #DCSR>, 47 %b: tensor<20x30xf32>, 48 %c: tensor<10x30xf32>) -> tensor<10x30xf32> { 49 %0 = linalg.matmul 50 ins(%a, %b: tensor<10x20xf32, #DCSR>, tensor<20x30xf32>) 51 outs(%c: tensor<10x30xf32>) -> tensor<10x30xf32> 52 return %0 : tensor<10x30xf32> 53} 54 55// 56// Computes C = A x B with all matrices sparse (SpMSpM) in DCSR. 57// 58// CHECK-LABEL: func @matmul2( 59// CHECK-SAME: %[[VAL_0:.*]]: tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>>, 60// CHECK-SAME: %[[VAL_1:.*]]: tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> { 61// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 0 : index 62// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 1 : index 63// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 2 : index 64// CHECK-DAG: %[[VAL_6:.*]] = arith.constant false 65// CHECK-DAG: %[[VAL_7:.*]] = arith.constant true 66// CHECK: %[[VAL_8:.*]] = bufferization.alloc_tensor() : tensor<4x4xf64, #sparse_tensor.encoding<{{{.*}}}>> 67// CHECK: %[[VAL_9:.*]] = sparse_tensor.pointers %[[VAL_0]], %[[VAL_3]] : tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 68// CHECK: %[[VAL_10:.*]] = sparse_tensor.indices %[[VAL_0]], %[[VAL_3]] : tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 69// CHECK: %[[VAL_11:.*]] = sparse_tensor.pointers %[[VAL_0]], %[[VAL_4]] : tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 70// CHECK: %[[VAL_12:.*]] = sparse_tensor.indices %[[VAL_0]], %[[VAL_4]] : tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 71// CHECK: %[[VAL_13:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xf64> 72// CHECK: %[[VAL_14:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_3]] : tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 73// CHECK: %[[VAL_15:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_3]] : tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 74// CHECK: %[[VAL_16:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_4]] : tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 75// CHECK: %[[VAL_17:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_4]] : tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 76// CHECK: %[[VAL_18:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xf64> 77// CHECK: %[[VAL_19:.*]] = memref.alloca(%[[VAL_5]]) : memref<?xindex> 78// CHECK: %[[VAL_20:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_3]]] : memref<?xindex> 79// CHECK: %[[VAL_21:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_4]]] : memref<?xindex> 80// CHECK: scf.for %[[VAL_22:.*]] = %[[VAL_20]] to %[[VAL_21]] step %[[VAL_4]] { 81// CHECK: %[[VAL_23:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_22]]] : memref<?xindex> 82// CHECK: memref.store %[[VAL_23]], %[[VAL_19]]{{\[}}%[[VAL_3]]] : memref<?xindex> 83// CHECK: %[[VAL_24:.*]], %[[VAL_25:.*]], %[[VAL_26:.*]], %[[VAL_27:.*]] = sparse_tensor.expand %[[VAL_8]] : tensor<4x4xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xf64>, memref<?xi1>, memref<?xindex>, index 84// CHECK: %[[VAL_28:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_22]]] : memref<?xindex> 85// CHECK: %[[VAL_29:.*]] = arith.addi %[[VAL_22]], %[[VAL_4]] : index 86// CHECK: %[[VAL_30:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_29]]] : memref<?xindex> 87// CHECK: %[[VAL_31:.*]] = memref.load %[[VAL_14]]{{\[}}%[[VAL_3]]] : memref<?xindex> 88// CHECK: %[[VAL_32:.*]] = memref.load %[[VAL_14]]{{\[}}%[[VAL_4]]] : memref<?xindex> 89// CHECK: %[[VAL_33:.*]]:3 = scf.while (%[[VAL_34:.*]] = %[[VAL_28]], %[[VAL_35:.*]] = %[[VAL_31]], %[[VAL_36:.*]] = %[[VAL_27]]) : (index, index, index) -> (index, index, index) { 90// CHECK: %[[VAL_37:.*]] = arith.cmpi ult, %[[VAL_34]], %[[VAL_30]] : index 91// CHECK: %[[VAL_38:.*]] = arith.cmpi ult, %[[VAL_35]], %[[VAL_32]] : index 92// CHECK: %[[VAL_39:.*]] = arith.andi %[[VAL_37]], %[[VAL_38]] : i1 93// CHECK: scf.condition(%[[VAL_39]]) %[[VAL_34]], %[[VAL_35]], %[[VAL_36]] : index, index, index 94// CHECK: } do { 95// CHECK: ^bb0(%[[VAL_40:.*]]: index, %[[VAL_41:.*]]: index, %[[VAL_42:.*]]: index): 96// CHECK: %[[VAL_43:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_40]]] : memref<?xindex> 97// CHECK: %[[VAL_44:.*]] = memref.load %[[VAL_15]]{{\[}}%[[VAL_41]]] : memref<?xindex> 98// CHECK: %[[VAL_45:.*]] = arith.cmpi ult, %[[VAL_44]], %[[VAL_43]] : index 99// CHECK: %[[VAL_46:.*]] = arith.select %[[VAL_45]], %[[VAL_44]], %[[VAL_43]] : index 100// CHECK: %[[VAL_47:.*]] = arith.cmpi eq, %[[VAL_43]], %[[VAL_46]] : index 101// CHECK: %[[VAL_48:.*]] = arith.cmpi eq, %[[VAL_44]], %[[VAL_46]] : index 102// CHECK: %[[VAL_49:.*]] = arith.andi %[[VAL_47]], %[[VAL_48]] : i1 103// CHECK: %[[VAL_50:.*]] = scf.if %[[VAL_49]] -> (index) { 104// CHECK: %[[VAL_51:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_40]]] : memref<?xf64> 105// CHECK: %[[VAL_52:.*]] = memref.load %[[VAL_16]]{{\[}}%[[VAL_41]]] : memref<?xindex> 106// CHECK: %[[VAL_53:.*]] = arith.addi %[[VAL_41]], %[[VAL_4]] : index 107// CHECK: %[[VAL_54:.*]] = memref.load %[[VAL_16]]{{\[}}%[[VAL_53]]] : memref<?xindex> 108// CHECK: %[[VAL_55:.*]] = scf.for %[[VAL_56:.*]] = %[[VAL_52]] to %[[VAL_54]] step %[[VAL_4]] iter_args(%[[VAL_57:.*]] = %[[VAL_42]]) -> (index) { 109// CHECK: %[[VAL_58:.*]] = memref.load %[[VAL_17]]{{\[}}%[[VAL_56]]] : memref<?xindex> 110// CHECK: %[[VAL_59:.*]] = memref.load %[[VAL_24]]{{\[}}%[[VAL_58]]] : memref<?xf64> 111// CHECK: %[[VAL_60:.*]] = memref.load %[[VAL_18]]{{\[}}%[[VAL_56]]] : memref<?xf64> 112// CHECK: %[[VAL_61:.*]] = arith.mulf %[[VAL_51]], %[[VAL_60]] : f64 113// CHECK: %[[VAL_62:.*]] = arith.addf %[[VAL_59]], %[[VAL_61]] : f64 114// CHECK: %[[VAL_63:.*]] = memref.load %[[VAL_25]]{{\[}}%[[VAL_58]]] : memref<?xi1> 115// CHECK: %[[VAL_64:.*]] = arith.cmpi eq, %[[VAL_63]], %[[VAL_6]] : i1 116// CHECK: %[[VAL_65:.*]] = scf.if %[[VAL_64]] -> (index) { 117// CHECK: memref.store %[[VAL_7]], %[[VAL_25]]{{\[}}%[[VAL_58]]] : memref<?xi1> 118// CHECK: memref.store %[[VAL_58]], %[[VAL_26]]{{\[}}%[[VAL_57]]] : memref<?xindex> 119// CHECK: %[[VAL_66:.*]] = arith.addi %[[VAL_57]], %[[VAL_4]] : index 120// CHECK: scf.yield %[[VAL_66]] : index 121// CHECK: } else { 122// CHECK: scf.yield %[[VAL_57]] : index 123// CHECK: } 124// CHECK: memref.store %[[VAL_62]], %[[VAL_24]]{{\[}}%[[VAL_58]]] : memref<?xf64> 125// CHECK: scf.yield %[[VAL_67:.*]] : index 126// CHECK: } 127// CHECK: scf.yield %[[VAL_68:.*]] : index 128// CHECK: } else { 129// CHECK: scf.yield %[[VAL_42]] : index 130// CHECK: } 131// CHECK: %[[VAL_69:.*]] = arith.cmpi eq, %[[VAL_43]], %[[VAL_46]] : index 132// CHECK: %[[VAL_70:.*]] = arith.addi %[[VAL_40]], %[[VAL_4]] : index 133// CHECK: %[[VAL_71:.*]] = arith.select %[[VAL_69]], %[[VAL_70]], %[[VAL_40]] : index 134// CHECK: %[[VAL_72:.*]] = arith.cmpi eq, %[[VAL_44]], %[[VAL_46]] : index 135// CHECK: %[[VAL_73:.*]] = arith.addi %[[VAL_41]], %[[VAL_4]] : index 136// CHECK: %[[VAL_74:.*]] = arith.select %[[VAL_72]], %[[VAL_73]], %[[VAL_41]] : index 137// CHECK: scf.yield %[[VAL_71]], %[[VAL_74]], %[[VAL_75:.*]] : index, index, index 138// CHECK: } 139// CHECK: sparse_tensor.compress %[[VAL_8]], %[[VAL_19]], %[[VAL_24]], %[[VAL_25]], %[[VAL_26]], %[[VAL_76:.*]]#2 : tensor<4x4xf64, #sparse_tensor.encoding<{{{.*}}}>>, memref<?xindex>, memref<?xf64>, memref<?xi1>, memref<?xindex>, index 140// CHECK: } 141// CHECK: %[[VAL_77:.*]] = sparse_tensor.load %[[VAL_8]] hasInserts : tensor<4x4xf64, #sparse_tensor.encoding<{{{.*}}}>> 142// CHECK: return %[[VAL_77]] : tensor<4x4xf64, #sparse_tensor.encoding<{{{.*}}}>> 143// CHECK: } 144func.func @matmul2(%A: tensor<4x8xf64, #DCSR>, 145 %B: tensor<8x4xf64, #DCSR>) -> tensor<4x4xf64, #DCSR> { 146 %c4 = arith.constant 4 : index 147 %C = bufferization.alloc_tensor() : tensor<4x4xf64, #DCSR> 148 %D = linalg.matmul 149 ins(%A, %B: tensor<4x8xf64, #DCSR>, tensor<8x4xf64, #DCSR>) 150 outs(%C: tensor<4x4xf64, #DCSR>) -> tensor<4x4xf64, #DCSR> 151 return %D: tensor<4x4xf64, #DCSR> 152} 153 154// CHECK-LABEL: func @conv2d( 155// CHECK-SAME: %[[VAL_0:.*]]: tensor<8x8xi32>, 156// CHECK-SAME: %[[VAL_1:.*]]: tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>>, 157// CHECK-SAME: %[[VAL_2:.*]]: tensor<6x6xi32>) -> tensor<6x6xi32> { 158// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 0 : index 159// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 1 : index 160// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 6 : index 161// CHECK-DAG: %[[VAL_6:.*]] = bufferization.to_memref %[[VAL_0]] : memref<8x8xi32> 162// CHECK-DAG: %[[VAL_7:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_3]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>> 163// CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_3]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>> 164// CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_4]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>> 165// CHECK-DAG: %[[VAL_10:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_4]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>> 166// CHECK-DAG: %[[VAL_11:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>> 167// CHECK-DAG: %[[VAL_13:.*]] = bufferization.to_memref %[[VAL_2]] : memref<6x6xi32> 168// CHECK: %[[VAL_14:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_3]]] : memref<?xindex> 169// CHECK: %[[VAL_15:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_4]]] : memref<?xindex> 170// CHECK: scf.for %[[VAL_16:.*]] = %[[VAL_14]] to %[[VAL_15]] step %[[VAL_4]] { 171// CHECK: %[[VAL_17:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_16]]] : memref<?xindex> 172// CHECK: %[[VAL_18:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_16]]] : memref<?xindex> 173// CHECK: %[[VAL_19:.*]] = arith.addi %[[VAL_16]], %[[VAL_4]] : index 174// CHECK: %[[VAL_20:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_19]]] : memref<?xindex> 175// CHECK: scf.for %[[VAL_21:.*]] = %[[VAL_18]] to %[[VAL_20]] step %[[VAL_4]] { 176// CHECK: %[[VAL_22:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_21]]] : memref<?xindex> 177// CHECK: %[[VAL_23:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_21]]] : memref<?xi32> 178// CHECK: scf.for %[[VAL_24:.*]] = %[[VAL_3]] to %[[VAL_5]] step %[[VAL_4]] { 179// CHECK: scf.for %[[VAL_25:.*]] = %[[VAL_3]] to %[[VAL_5]] step %[[VAL_4]] { 180// CHECK: %[[VAL_26:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_25]], %[[VAL_24]]] : memref<6x6xi32> 181// CHECK: %[[VAL_27:.*]] = arith.addi %[[VAL_25]], %[[VAL_17]] : index 182// CHECK: %[[VAL_28:.*]] = arith.addi %[[VAL_24]], %[[VAL_22]] : index 183// CHECK: %[[VAL_29:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_27]], %[[VAL_28]]] : memref<8x8xi32> 184// CHECK: %[[VAL_30:.*]] = arith.muli %[[VAL_29]], %[[VAL_23]] : i32 185// CHECK: %[[VAL_31:.*]] = arith.addi %[[VAL_26]], %[[VAL_30]] : i32 186// CHECK: memref.store %[[VAL_31]], %[[VAL_13]]{{\[}}%[[VAL_25]], %[[VAL_24]]] : memref<6x6xi32> 187// CHECK: } 188// CHECK: } 189// CHECK: } 190// CHECK: } 191// CHECK: %[[VAL_32:.*]] = bufferization.to_tensor %[[VAL_13]] : memref<6x6xi32> 192// CHECK: return %[[VAL_32]] : tensor<6x6xi32> 193// CHECK: } 194func.func @conv2d(%input: tensor<8x8xi32>, 195 %filter: tensor<3x3xi32, #DCSR>, 196 %output: tensor<6x6xi32>) -> tensor<6x6xi32> { 197 %0 = linalg.conv_2d 198 ins (%input, %filter: tensor<8x8xi32>, tensor<3x3xi32, #DCSR>) 199 outs (%output: tensor<6x6xi32>) -> tensor<6x6xi32> 200 return %0 : tensor<6x6xi32> 201} 202 203// CHECK-LABEL: func @quantized_matmul( 204// CHECK-SAME: %[[VAL_0:.*]]: tensor<5x3xi8>, 205// CHECK-SAME: %[[VAL_1:.*]]: tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>>, 206// CHECK-SAME: %[[VAL_2:.*]]: tensor<5x6xi64>) -> tensor<5x6xi64> { 207// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 2 : i64 208// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 0 : index 209// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 1 : index 210// CHECK-DAG: %[[VAL_6:.*]] = arith.constant 5 : index 211// CHECK-DAG: %[[VAL_7:.*]] = bufferization.to_memref %[[VAL_0]] : memref<5x3xi8> 212// CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_4]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>> 213// CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_4]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>> 214// CHECK-DAG: %[[VAL_10:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_5]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>> 215// CHECK-DAG: %[[VAL_11:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_5]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>> 216// CHECK-DAG: %[[VAL_12:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>> 217// CHECK-DAG: %[[VAL_14:.*]] = bufferization.to_memref %[[VAL_2]] : memref<5x6xi64> 218// CHECK: %[[VAL_15:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex> 219// CHECK: %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_5]]] : memref<?xindex> 220// CHECK: scf.for %[[VAL_17:.*]] = %[[VAL_15]] to %[[VAL_16]] step %[[VAL_5]] { 221// CHECK: %[[VAL_18:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_17]]] : memref<?xindex> 222// CHECK: %[[VAL_19:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_17]]] : memref<?xindex> 223// CHECK: %[[VAL_20:.*]] = arith.addi %[[VAL_17]], %[[VAL_5]] : index 224// CHECK: %[[VAL_21:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_20]]] : memref<?xindex> 225// CHECK: scf.for %[[VAL_22:.*]] = %[[VAL_19]] to %[[VAL_21]] step %[[VAL_5]] { 226// CHECK: %[[VAL_23:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_22]]] : memref<?xindex> 227// CHECK: %[[VAL_24:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_22]]] : memref<?xi8> 228// CHECK: scf.for %[[VAL_25:.*]] = %[[VAL_4]] to %[[VAL_6]] step %[[VAL_5]] { 229// CHECK: %[[VAL_26:.*]] = memref.load %[[VAL_14]]{{\[}}%[[VAL_25]], %[[VAL_23]]] : memref<5x6xi64> 230// CHECK: %[[VAL_27:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_25]], %[[VAL_18]]] : memref<5x3xi8> 231// CHECK: %[[VAL_28:.*]] = arith.extsi %[[VAL_27]] : i8 to i64 232// CHECK: %[[VAL_29:.*]] = arith.subi %[[VAL_28]], %[[VAL_3]] : i64 233// CHECK: %[[VAL_30:.*]] = arith.extsi %[[VAL_24]] : i8 to i64 234// CHECK: %[[VAL_31:.*]] = arith.muli %[[VAL_29]], %[[VAL_30]] : i64 235// CHECK: %[[VAL_32:.*]] = arith.addi %[[VAL_26]], %[[VAL_31]] : i64 236// CHECK: memref.store %[[VAL_32]], %[[VAL_14]]{{\[}}%[[VAL_25]], %[[VAL_23]]] : memref<5x6xi64> 237// CHECK: } 238// CHECK: } 239// CHECK: } 240// CHECK: %[[VAL_33:.*]] = bufferization.to_tensor %[[VAL_14]] : memref<5x6xi64> 241// CHECK: return %[[VAL_33]] : tensor<5x6xi64> 242// CHECK: } 243func.func @quantized_matmul(%input1: tensor<5x3xi8>, 244 %input2: tensor<3x6xi8, #DCSR>, 245 %output: tensor<5x6xi64>) -> tensor<5x6xi64> { 246 %c0 = arith.constant 0 : i32 247 %c2 = arith.constant 2 : i32 248 %0 = linalg.quantized_matmul 249 ins(%input1, %input2, %c2, %c0 : tensor<5x3xi8>, tensor<3x6xi8, #DCSR>, i32, i32) 250 outs(%output : tensor<5x6xi64>) -> tensor<5x6xi64> 251 return %0: tensor<5x6xi64> 252} 253 254// CHECK-LABEL: func @sparse_dot( 255// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 0 : index 256// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 1 : index 257// CHECK-DAG: %[[VAL_5:.*]] = sparse_tensor.pointers %[[VAL_0:.*]], %[[VAL_3]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 258// CHECK-DAG: %[[VAL_6:.*]] = sparse_tensor.indices %[[VAL_0]], %[[VAL_3]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 259// CHECK-DAG: %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xf32> 260// CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_1:.*]], %[[VAL_3]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 261// CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_3]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 262// CHECK-DAG: %[[VAL_10:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xf32> 263// CHECK-DAG: %[[VAL_11:.*]] = bufferization.to_memref %[[VAL_2:.*]] : memref<f32> 264// CHECK-DAG: %[[VAL_13:.*]] = memref.load %[[VAL_11]][] : memref<f32> 265// CHECK-DAG: %[[VAL_14:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex> 266// CHECK-DAG: %[[VAL_15:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex> 267// CHECK-DAG: %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_3]]] : memref<?xindex> 268// CHECK-DAG: %[[VAL_17:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex> 269// CHECK: %[[VAL_18:.*]]:3 = scf.while (%[[VAL_19:.*]] = %[[VAL_14]], %[[VAL_20:.*]] = %[[VAL_16]], %[[VAL_21:.*]] = %[[VAL_13]]) : (index, index, f32) -> (index, index, f32) { 270// CHECK: %[[VAL_22:.*]] = arith.cmpi ult, %[[VAL_19]], %[[VAL_15]] : index 271// CHECK: %[[VAL_23:.*]] = arith.cmpi ult, %[[VAL_20]], %[[VAL_17]] : index 272// CHECK: %[[VAL_24:.*]] = arith.andi %[[VAL_22]], %[[VAL_23]] : i1 273// CHECK: scf.condition(%[[VAL_24]]) %[[VAL_19]], %[[VAL_20]], %[[VAL_21]] : index, index, f32 274// CHECK: } do { 275// CHECK: ^bb0(%[[VAL_25:.*]]: index, %[[VAL_26:.*]]: index, %[[VAL_27:.*]]: f32): 276// CHECK: %[[VAL_28:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_25]]] : memref<?xindex> 277// CHECK: %[[VAL_29:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_26]]] : memref<?xindex> 278// CHECK: %[[VAL_30:.*]] = arith.cmpi ult, %[[VAL_29]], %[[VAL_28]] : index 279// CHECK: %[[VAL_31:.*]] = arith.select %[[VAL_30]], %[[VAL_29]], %[[VAL_28]] : index 280// CHECK: %[[VAL_32:.*]] = arith.cmpi eq, %[[VAL_28]], %[[VAL_31]] : index 281// CHECK: %[[VAL_33:.*]] = arith.cmpi eq, %[[VAL_29]], %[[VAL_31]] : index 282// CHECK: %[[VAL_34:.*]] = arith.andi %[[VAL_32]], %[[VAL_33]] : i1 283// CHECK: %[[VAL_35:.*]] = scf.if %[[VAL_34]] -> (f32) { 284// CHECK: %[[VAL_36:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_25]]] : memref<?xf32> 285// CHECK: %[[VAL_37:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_26]]] : memref<?xf32> 286// CHECK: %[[VAL_38:.*]] = arith.mulf %[[VAL_36]], %[[VAL_37]] : f32 287// CHECK: %[[VAL_39:.*]] = arith.addf %[[VAL_27]], %[[VAL_38]] : f32 288// CHECK: scf.yield %[[VAL_39]] : f32 289// CHECK: } else { 290// CHECK: scf.yield %[[VAL_27]] : f32 291// CHECK: } 292// CHECK: %[[VAL_40:.*]] = arith.cmpi eq, %[[VAL_28]], %[[VAL_31]] : index 293// CHECK: %[[VAL_41:.*]] = arith.addi %[[VAL_25]], %[[VAL_4]] : index 294// CHECK: %[[VAL_42:.*]] = arith.select %[[VAL_40]], %[[VAL_41]], %[[VAL_25]] : index 295// CHECK: %[[VAL_43:.*]] = arith.cmpi eq, %[[VAL_29]], %[[VAL_31]] : index 296// CHECK: %[[VAL_44:.*]] = arith.addi %[[VAL_26]], %[[VAL_4]] : index 297// CHECK: %[[VAL_45:.*]] = arith.select %[[VAL_43]], %[[VAL_44]], %[[VAL_26]] : index 298// CHECK: scf.yield %[[VAL_42]], %[[VAL_45]], %[[VAL_46:.*]] : index, index, f32 299// CHECK: } 300// CHECK: memref.store %[[VAL_47:.*]]#2, %[[VAL_11]][] : memref<f32> 301// CHECK: %[[VAL_48:.*]] = bufferization.to_tensor %[[VAL_11]] : memref<f32> 302// CHECK: return %[[VAL_48]] : tensor<f32> 303// CHECK: } 304func.func @sparse_dot(%a: tensor<1024xf32, #SparseVector>, 305 %b: tensor<1024xf32, #SparseVector>, 306 %x: tensor<f32>) -> tensor<f32> { 307 %dot = linalg.dot ins(%a, %b: tensor<1024xf32, #SparseVector>, 308 tensor<1024xf32, #SparseVector>) 309 outs(%x: tensor<f32>) -> tensor<f32> 310 return %dot : tensor<f32> 311} 312