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_12:.*]] = bufferization.to_memref %[[VAL_2]] : memref<10x30xf32> 24// CHECK-DAG: %[[VAL_13:.*]] = memref.alloc() : memref<10x30xf32> 25// CHECK: memref.copy %[[VAL_12]], %[[VAL_13]] : memref<10x30xf32> to memref<10x30xf32> 26// CHECK: %[[VAL_14:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_3]]] : memref<?xindex> 27// CHECK: %[[VAL_15:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_4]]] : memref<?xindex> 28// CHECK: scf.for %[[VAL_16:.*]] = %[[VAL_14]] to %[[VAL_15]] step %[[VAL_4]] { 29// CHECK: %[[VAL_17:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_16]]] : memref<?xindex> 30// CHECK: %[[VAL_18:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_16]]] : memref<?xindex> 31// CHECK: %[[VAL_19:.*]] = arith.addi %[[VAL_16]], %[[VAL_4]] : index 32// CHECK: %[[VAL_20:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_19]]] : memref<?xindex> 33// CHECK: scf.for %[[VAL_21:.*]] = %[[VAL_18]] to %[[VAL_20]] step %[[VAL_4]] { 34// CHECK: %[[VAL_22:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_21]]] : memref<?xindex> 35// CHECK: %[[VAL_23:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_21]]] : memref<?xf32> 36// CHECK: scf.for %[[VAL_24:.*]] = %[[VAL_3]] to %[[VAL_5]] step %[[VAL_4]] { 37// CHECK: %[[VAL_25:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_17]], %[[VAL_24]]] : memref<10x30xf32> 38// CHECK: %[[VAL_26:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_22]], %[[VAL_24]]] : memref<20x30xf32> 39// CHECK: %[[VAL_27:.*]] = arith.mulf %[[VAL_23]], %[[VAL_26]] : f32 40// CHECK: %[[VAL_28:.*]] = arith.addf %[[VAL_25]], %[[VAL_27]] : f32 41// CHECK: memref.store %[[VAL_28]], %[[VAL_13]]{{\[}}%[[VAL_17]], %[[VAL_24]]] : memref<10x30xf32> 42// CHECK: } 43// CHECK: } 44// CHECK: } 45// CHECK: %[[VAL_29:.*]] = bufferization.to_tensor %[[VAL_13]] : memref<10x30xf32> 46// CHECK: return %[[VAL_29]] : tensor<10x30xf32> 47// CHECK: } 48func.func @matmul1(%a: tensor<10x20xf32, #DCSR>, 49 %b: tensor<20x30xf32>, 50 %c: tensor<10x30xf32>) -> tensor<10x30xf32> { 51 %0 = linalg.matmul 52 ins(%a, %b: tensor<10x20xf32, #DCSR>, tensor<20x30xf32>) 53 outs(%c: tensor<10x30xf32>) -> tensor<10x30xf32> 54 return %0 : tensor<10x30xf32> 55} 56 57// 58// Computes C = A x B with all matrices sparse (SpMSpM) in DCSR. 59// 60// CHECK-LABEL: func @matmul2( 61// CHECK-SAME: %[[VAL_0:.*]]: tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>>, 62// CHECK-SAME: %[[VAL_1:.*]]: tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> { 63// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 0 : index 64// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 1 : index 65// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 2 : index 66// CHECK-DAG: %[[VAL_6:.*]] = arith.constant false 67// CHECK-DAG: %[[VAL_7:.*]] = arith.constant true 68// CHECK: %[[VAL_8:.*]] = bufferization.alloc_tensor() : tensor<4x4xf64, #sparse_tensor.encoding<{{{.*}}}>> 69// CHECK: %[[VAL_9:.*]] = sparse_tensor.pointers %[[VAL_0]], %[[VAL_3]] : tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 70// CHECK: %[[VAL_10:.*]] = sparse_tensor.indices %[[VAL_0]], %[[VAL_3]] : tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 71// CHECK: %[[VAL_11:.*]] = sparse_tensor.pointers %[[VAL_0]], %[[VAL_4]] : tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 72// CHECK: %[[VAL_12:.*]] = sparse_tensor.indices %[[VAL_0]], %[[VAL_4]] : tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 73// CHECK: %[[VAL_13:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<4x8xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xf64> 74// CHECK: %[[VAL_14:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_3]] : tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 75// CHECK: %[[VAL_15:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_3]] : tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 76// CHECK: %[[VAL_16:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_4]] : tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 77// CHECK: %[[VAL_17:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_4]] : tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 78// CHECK: %[[VAL_18:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<8x4xf64, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xf64> 79// CHECK: %[[VAL_19:.*]] = memref.alloca(%[[VAL_5]]) : memref<?xindex> 80// CHECK: %[[VAL_20:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_3]]] : memref<?xindex> 81// CHECK: %[[VAL_21:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_4]]] : memref<?xindex> 82// CHECK: scf.for %[[VAL_22:.*]] = %[[VAL_20]] to %[[VAL_21]] step %[[VAL_4]] { 83// CHECK: %[[VAL_23:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_22]]] : memref<?xindex> 84// CHECK: memref.store %[[VAL_23]], %[[VAL_19]]{{\[}}%[[VAL_3]]] : memref<?xindex> 85// 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 86// CHECK: %[[VAL_28:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_22]]] : memref<?xindex> 87// CHECK: %[[VAL_29:.*]] = arith.addi %[[VAL_22]], %[[VAL_4]] : index 88// CHECK: %[[VAL_30:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_29]]] : memref<?xindex> 89// CHECK: %[[VAL_31:.*]] = memref.load %[[VAL_14]]{{\[}}%[[VAL_3]]] : memref<?xindex> 90// CHECK: %[[VAL_32:.*]] = memref.load %[[VAL_14]]{{\[}}%[[VAL_4]]] : memref<?xindex> 91// CHECK: %[[VAL_33:.*]]:3 = scf.while (%[[VAL_34:.*]] = %[[VAL_28]], %[[VAL_35:.*]] = %[[VAL_31]], %[[VAL_36:.*]] = %[[VAL_27]]) : (index, index, index) -> (index, index, index) { 92// CHECK: %[[VAL_37:.*]] = arith.cmpi ult, %[[VAL_34]], %[[VAL_30]] : index 93// CHECK: %[[VAL_38:.*]] = arith.cmpi ult, %[[VAL_35]], %[[VAL_32]] : index 94// CHECK: %[[VAL_39:.*]] = arith.andi %[[VAL_37]], %[[VAL_38]] : i1 95// CHECK: scf.condition(%[[VAL_39]]) %[[VAL_34]], %[[VAL_35]], %[[VAL_36]] : index, index, index 96// CHECK: } do { 97// CHECK: ^bb0(%[[VAL_40:.*]]: index, %[[VAL_41:.*]]: index, %[[VAL_42:.*]]: index): 98// CHECK: %[[VAL_43:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_40]]] : memref<?xindex> 99// CHECK: %[[VAL_44:.*]] = memref.load %[[VAL_15]]{{\[}}%[[VAL_41]]] : memref<?xindex> 100// CHECK: %[[VAL_45:.*]] = arith.cmpi ult, %[[VAL_44]], %[[VAL_43]] : index 101// CHECK: %[[VAL_46:.*]] = arith.select %[[VAL_45]], %[[VAL_44]], %[[VAL_43]] : index 102// CHECK: %[[VAL_47:.*]] = arith.cmpi eq, %[[VAL_43]], %[[VAL_46]] : index 103// CHECK: %[[VAL_48:.*]] = arith.cmpi eq, %[[VAL_44]], %[[VAL_46]] : index 104// CHECK: %[[VAL_49:.*]] = arith.andi %[[VAL_47]], %[[VAL_48]] : i1 105// CHECK: %[[VAL_50:.*]] = scf.if %[[VAL_49]] -> (index) { 106// CHECK: %[[VAL_51:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_40]]] : memref<?xf64> 107// CHECK: %[[VAL_52:.*]] = memref.load %[[VAL_16]]{{\[}}%[[VAL_41]]] : memref<?xindex> 108// CHECK: %[[VAL_53:.*]] = arith.addi %[[VAL_41]], %[[VAL_4]] : index 109// CHECK: %[[VAL_54:.*]] = memref.load %[[VAL_16]]{{\[}}%[[VAL_53]]] : memref<?xindex> 110// CHECK: %[[VAL_55:.*]] = scf.for %[[VAL_56:.*]] = %[[VAL_52]] to %[[VAL_54]] step %[[VAL_4]] iter_args(%[[VAL_57:.*]] = %[[VAL_42]]) -> (index) { 111// CHECK: %[[VAL_58:.*]] = memref.load %[[VAL_17]]{{\[}}%[[VAL_56]]] : memref<?xindex> 112// CHECK: %[[VAL_59:.*]] = memref.load %[[VAL_24]]{{\[}}%[[VAL_58]]] : memref<?xf64> 113// CHECK: %[[VAL_60:.*]] = memref.load %[[VAL_18]]{{\[}}%[[VAL_56]]] : memref<?xf64> 114// CHECK: %[[VAL_61:.*]] = arith.mulf %[[VAL_51]], %[[VAL_60]] : f64 115// CHECK: %[[VAL_62:.*]] = arith.addf %[[VAL_59]], %[[VAL_61]] : f64 116// CHECK: %[[VAL_63:.*]] = memref.load %[[VAL_25]]{{\[}}%[[VAL_58]]] : memref<?xi1> 117// CHECK: %[[VAL_64:.*]] = arith.cmpi eq, %[[VAL_63]], %[[VAL_6]] : i1 118// CHECK: %[[VAL_65:.*]] = scf.if %[[VAL_64]] -> (index) { 119// CHECK: memref.store %[[VAL_7]], %[[VAL_25]]{{\[}}%[[VAL_58]]] : memref<?xi1> 120// CHECK: memref.store %[[VAL_58]], %[[VAL_26]]{{\[}}%[[VAL_57]]] : memref<?xindex> 121// CHECK: %[[VAL_66:.*]] = arith.addi %[[VAL_57]], %[[VAL_4]] : index 122// CHECK: scf.yield %[[VAL_66]] : index 123// CHECK: } else { 124// CHECK: scf.yield %[[VAL_57]] : index 125// CHECK: } 126// CHECK: memref.store %[[VAL_62]], %[[VAL_24]]{{\[}}%[[VAL_58]]] : memref<?xf64> 127// CHECK: scf.yield %[[VAL_67:.*]] : index 128// CHECK: } 129// CHECK: scf.yield %[[VAL_68:.*]] : index 130// CHECK: } else { 131// CHECK: scf.yield %[[VAL_42]] : index 132// CHECK: } 133// CHECK: %[[VAL_69:.*]] = arith.cmpi eq, %[[VAL_43]], %[[VAL_46]] : index 134// CHECK: %[[VAL_70:.*]] = arith.addi %[[VAL_40]], %[[VAL_4]] : index 135// CHECK: %[[VAL_71:.*]] = arith.select %[[VAL_69]], %[[VAL_70]], %[[VAL_40]] : index 136// CHECK: %[[VAL_72:.*]] = arith.cmpi eq, %[[VAL_44]], %[[VAL_46]] : index 137// CHECK: %[[VAL_73:.*]] = arith.addi %[[VAL_41]], %[[VAL_4]] : index 138// CHECK: %[[VAL_74:.*]] = arith.select %[[VAL_72]], %[[VAL_73]], %[[VAL_41]] : index 139// CHECK: scf.yield %[[VAL_71]], %[[VAL_74]], %[[VAL_75:.*]] : index, index, index 140// CHECK: } 141// 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 142// CHECK: } 143// CHECK: %[[VAL_77:.*]] = sparse_tensor.load %[[VAL_8]] hasInserts : tensor<4x4xf64, #sparse_tensor.encoding<{{{.*}}}>> 144// CHECK: return %[[VAL_77]] : tensor<4x4xf64, #sparse_tensor.encoding<{{{.*}}}>> 145// CHECK: } 146func.func @matmul2(%A: tensor<4x8xf64, #DCSR>, 147 %B: tensor<8x4xf64, #DCSR>) -> tensor<4x4xf64, #DCSR> { 148 %c4 = arith.constant 4 : index 149 %C = bufferization.alloc_tensor() : tensor<4x4xf64, #DCSR> 150 %D = linalg.matmul 151 ins(%A, %B: tensor<4x8xf64, #DCSR>, tensor<8x4xf64, #DCSR>) 152 outs(%C: tensor<4x4xf64, #DCSR>) -> tensor<4x4xf64, #DCSR> 153 return %D: tensor<4x4xf64, #DCSR> 154} 155 156// CHECK-LABEL: func @conv2d( 157// CHECK-SAME: %[[VAL_0:.*]]: tensor<8x8xi32>, 158// CHECK-SAME: %[[VAL_1:.*]]: tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>>, 159// CHECK-SAME: %[[VAL_2:.*]]: tensor<6x6xi32>) -> tensor<6x6xi32> { 160// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 0 : index 161// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 1 : index 162// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 6 : index 163// CHECK-DAG: %[[VAL_6:.*]] = bufferization.to_memref %[[VAL_0]] : memref<8x8xi32> 164// CHECK-DAG: %[[VAL_7:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_3]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>> 165// CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_3]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>> 166// CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_4]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>> 167// CHECK-DAG: %[[VAL_10:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_4]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>> 168// CHECK-DAG: %[[VAL_11:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>> 169// CHECK-DAG: %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_2]] : memref<6x6xi32> 170// CHECK-DAG: %[[VAL_13:.*]] = memref.alloc() : memref<6x6xi32> 171// CHECK: memref.copy %[[VAL_12]], %[[VAL_13]] : memref<6x6xi32> to memref<6x6xi32> 172// CHECK: %[[VAL_14:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_3]]] : memref<?xindex> 173// CHECK: %[[VAL_15:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_4]]] : memref<?xindex> 174// CHECK: scf.for %[[VAL_16:.*]] = %[[VAL_14]] to %[[VAL_15]] step %[[VAL_4]] { 175// CHECK: %[[VAL_17:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_16]]] : memref<?xindex> 176// CHECK: %[[VAL_18:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_16]]] : memref<?xindex> 177// CHECK: %[[VAL_19:.*]] = arith.addi %[[VAL_16]], %[[VAL_4]] : index 178// CHECK: %[[VAL_20:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_19]]] : memref<?xindex> 179// CHECK: scf.for %[[VAL_21:.*]] = %[[VAL_18]] to %[[VAL_20]] step %[[VAL_4]] { 180// CHECK: %[[VAL_22:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_21]]] : memref<?xindex> 181// CHECK: %[[VAL_23:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_21]]] : memref<?xi32> 182// CHECK: scf.for %[[VAL_24:.*]] = %[[VAL_3]] to %[[VAL_5]] step %[[VAL_4]] { 183// CHECK: scf.for %[[VAL_25:.*]] = %[[VAL_3]] to %[[VAL_5]] step %[[VAL_4]] { 184// CHECK: %[[VAL_26:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_25]], %[[VAL_24]]] : memref<6x6xi32> 185// CHECK: %[[VAL_27:.*]] = arith.addi %[[VAL_25]], %[[VAL_17]] : index 186// CHECK: %[[VAL_28:.*]] = arith.addi %[[VAL_24]], %[[VAL_22]] : index 187// CHECK: %[[VAL_29:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_27]], %[[VAL_28]]] : memref<8x8xi32> 188// CHECK: %[[VAL_30:.*]] = arith.muli %[[VAL_29]], %[[VAL_23]] : i32 189// CHECK: %[[VAL_31:.*]] = arith.addi %[[VAL_26]], %[[VAL_30]] : i32 190// CHECK: memref.store %[[VAL_31]], %[[VAL_13]]{{\[}}%[[VAL_25]], %[[VAL_24]]] : memref<6x6xi32> 191// CHECK: } 192// CHECK: } 193// CHECK: } 194// CHECK: } 195// CHECK: %[[VAL_32:.*]] = bufferization.to_tensor %[[VAL_13]] : memref<6x6xi32> 196// CHECK: return %[[VAL_32]] : tensor<6x6xi32> 197// CHECK: } 198func.func @conv2d(%input: tensor<8x8xi32>, 199 %filter: tensor<3x3xi32, #DCSR>, 200 %output: tensor<6x6xi32>) -> tensor<6x6xi32> { 201 %0 = linalg.conv_2d 202 ins (%input, %filter: tensor<8x8xi32>, tensor<3x3xi32, #DCSR>) 203 outs (%output: tensor<6x6xi32>) -> tensor<6x6xi32> 204 return %0 : tensor<6x6xi32> 205} 206 207// CHECK-LABEL: func @quantized_matmul( 208// CHECK-SAME: %[[VAL_0:.*]]: tensor<5x3xi8>, 209// CHECK-SAME: %[[VAL_1:.*]]: tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>>, 210// CHECK-SAME: %[[VAL_2:.*]]: tensor<5x6xi64>) -> tensor<5x6xi64> { 211// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 2 : i64 212// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 0 : index 213// CHECK-DAG: %[[VAL_5:.*]] = arith.constant 1 : index 214// CHECK-DAG: %[[VAL_6:.*]] = arith.constant 5 : index 215// CHECK-DAG: %[[VAL_7:.*]] = bufferization.to_memref %[[VAL_0]] : memref<5x3xi8> 216// CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_4]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>> 217// CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_4]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>> 218// CHECK-DAG: %[[VAL_10:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_5]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>> 219// CHECK-DAG: %[[VAL_11:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_5]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>> 220// CHECK-DAG: %[[VAL_12:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>> 221// CHECK-DAG: %[[VAL_13:.*]] = bufferization.to_memref %[[VAL_2]] : memref<5x6xi64> 222// CHECK-DAG: %[[VAL_14:.*]] = memref.alloc() : memref<5x6xi64> 223// CHECK: memref.copy %[[VAL_13]], %[[VAL_14]] : memref<5x6xi64> to memref<5x6xi64> 224// CHECK: %[[VAL_15:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex> 225// CHECK: %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_5]]] : memref<?xindex> 226// CHECK: scf.for %[[VAL_17:.*]] = %[[VAL_15]] to %[[VAL_16]] step %[[VAL_5]] { 227// CHECK: %[[VAL_18:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_17]]] : memref<?xindex> 228// CHECK: %[[VAL_19:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_17]]] : memref<?xindex> 229// CHECK: %[[VAL_20:.*]] = arith.addi %[[VAL_17]], %[[VAL_5]] : index 230// CHECK: %[[VAL_21:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_20]]] : memref<?xindex> 231// CHECK: scf.for %[[VAL_22:.*]] = %[[VAL_19]] to %[[VAL_21]] step %[[VAL_5]] { 232// CHECK: %[[VAL_23:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_22]]] : memref<?xindex> 233// CHECK: %[[VAL_24:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_22]]] : memref<?xi8> 234// CHECK: scf.for %[[VAL_25:.*]] = %[[VAL_4]] to %[[VAL_6]] step %[[VAL_5]] { 235// CHECK: %[[VAL_26:.*]] = memref.load %[[VAL_14]]{{\[}}%[[VAL_25]], %[[VAL_23]]] : memref<5x6xi64> 236// CHECK: %[[VAL_27:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_25]], %[[VAL_18]]] : memref<5x3xi8> 237// CHECK: %[[VAL_28:.*]] = arith.extsi %[[VAL_27]] : i8 to i64 238// CHECK: %[[VAL_29:.*]] = arith.subi %[[VAL_28]], %[[VAL_3]] : i64 239// CHECK: %[[VAL_30:.*]] = arith.extsi %[[VAL_24]] : i8 to i64 240// CHECK: %[[VAL_31:.*]] = arith.muli %[[VAL_29]], %[[VAL_30]] : i64 241// CHECK: %[[VAL_32:.*]] = arith.addi %[[VAL_26]], %[[VAL_31]] : i64 242// CHECK: memref.store %[[VAL_32]], %[[VAL_14]]{{\[}}%[[VAL_25]], %[[VAL_23]]] : memref<5x6xi64> 243// CHECK: } 244// CHECK: } 245// CHECK: } 246// CHECK: %[[VAL_33:.*]] = bufferization.to_tensor %[[VAL_14]] : memref<5x6xi64> 247// CHECK: return %[[VAL_33]] : tensor<5x6xi64> 248// CHECK: } 249func.func @quantized_matmul(%input1: tensor<5x3xi8>, 250 %input2: tensor<3x6xi8, #DCSR>, 251 %output: tensor<5x6xi64>) -> tensor<5x6xi64> { 252 %c0 = arith.constant 0 : i32 253 %c2 = arith.constant 2 : i32 254 %0 = linalg.quantized_matmul 255 ins(%input1, %input2, %c2, %c0 : tensor<5x3xi8>, tensor<3x6xi8, #DCSR>, i32, i32) 256 outs(%output : tensor<5x6xi64>) -> tensor<5x6xi64> 257 return %0: tensor<5x6xi64> 258} 259 260// CHECK-LABEL: func @sparse_dot( 261// CHECK-DAG: %[[VAL_3:.*]] = arith.constant 0 : index 262// CHECK-DAG: %[[VAL_4:.*]] = arith.constant 1 : index 263// CHECK-DAG: %[[VAL_5:.*]] = sparse_tensor.pointers %[[VAL_0:.*]], %[[VAL_3]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 264// CHECK-DAG: %[[VAL_6:.*]] = sparse_tensor.indices %[[VAL_0]], %[[VAL_3]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 265// CHECK-DAG: %[[VAL_7:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xf32> 266// CHECK-DAG: %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_1:.*]], %[[VAL_3]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 267// CHECK-DAG: %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_3]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xindex> 268// CHECK-DAG: %[[VAL_10:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<1024xf32, #sparse_tensor.encoding<{{{.*}}}>> to memref<?xf32> 269// CHECK-DAG: %[[VAL_11:.*]] = memref.alloc() : memref<f32> 270// CHECK-DAG: %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_2:.*]] : memref<f32> 271// CHECK-DAG: memref.copy %[[VAL_12]], %[[VAL_11]] : memref<f32> to memref<f32> 272// CHECK-DAG: %[[VAL_13:.*]] = memref.load %[[VAL_11]][] : memref<f32> 273// CHECK-DAG: %[[VAL_14:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_3]]] : memref<?xindex> 274// CHECK-DAG: %[[VAL_15:.*]] = memref.load %[[VAL_5]]{{\[}}%[[VAL_4]]] : memref<?xindex> 275// CHECK-DAG: %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_3]]] : memref<?xindex> 276// CHECK-DAG: %[[VAL_17:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex> 277// CHECK: %[[VAL_18:.*]]:3 = scf.while (%[[VAL_19:.*]] = %[[VAL_14]], %[[VAL_20:.*]] = %[[VAL_16]], %[[VAL_21:.*]] = %[[VAL_13]]) : (index, index, f32) -> (index, index, f32) { 278// CHECK: %[[VAL_22:.*]] = arith.cmpi ult, %[[VAL_19]], %[[VAL_15]] : index 279// CHECK: %[[VAL_23:.*]] = arith.cmpi ult, %[[VAL_20]], %[[VAL_17]] : index 280// CHECK: %[[VAL_24:.*]] = arith.andi %[[VAL_22]], %[[VAL_23]] : i1 281// CHECK: scf.condition(%[[VAL_24]]) %[[VAL_19]], %[[VAL_20]], %[[VAL_21]] : index, index, f32 282// CHECK: } do { 283// CHECK: ^bb0(%[[VAL_25:.*]]: index, %[[VAL_26:.*]]: index, %[[VAL_27:.*]]: f32): 284// CHECK: %[[VAL_28:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_25]]] : memref<?xindex> 285// CHECK: %[[VAL_29:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_26]]] : memref<?xindex> 286// CHECK: %[[VAL_30:.*]] = arith.cmpi ult, %[[VAL_29]], %[[VAL_28]] : index 287// CHECK: %[[VAL_31:.*]] = arith.select %[[VAL_30]], %[[VAL_29]], %[[VAL_28]] : index 288// CHECK: %[[VAL_32:.*]] = arith.cmpi eq, %[[VAL_28]], %[[VAL_31]] : index 289// CHECK: %[[VAL_33:.*]] = arith.cmpi eq, %[[VAL_29]], %[[VAL_31]] : index 290// CHECK: %[[VAL_34:.*]] = arith.andi %[[VAL_32]], %[[VAL_33]] : i1 291// CHECK: %[[VAL_35:.*]] = scf.if %[[VAL_34]] -> (f32) { 292// CHECK: %[[VAL_36:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_25]]] : memref<?xf32> 293// CHECK: %[[VAL_37:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_26]]] : memref<?xf32> 294// CHECK: %[[VAL_38:.*]] = arith.mulf %[[VAL_36]], %[[VAL_37]] : f32 295// CHECK: %[[VAL_39:.*]] = arith.addf %[[VAL_27]], %[[VAL_38]] : f32 296// CHECK: scf.yield %[[VAL_39]] : f32 297// CHECK: } else { 298// CHECK: scf.yield %[[VAL_27]] : f32 299// CHECK: } 300// CHECK: %[[VAL_40:.*]] = arith.cmpi eq, %[[VAL_28]], %[[VAL_31]] : index 301// CHECK: %[[VAL_41:.*]] = arith.addi %[[VAL_25]], %[[VAL_4]] : index 302// CHECK: %[[VAL_42:.*]] = arith.select %[[VAL_40]], %[[VAL_41]], %[[VAL_25]] : index 303// CHECK: %[[VAL_43:.*]] = arith.cmpi eq, %[[VAL_29]], %[[VAL_31]] : index 304// CHECK: %[[VAL_44:.*]] = arith.addi %[[VAL_26]], %[[VAL_4]] : index 305// CHECK: %[[VAL_45:.*]] = arith.select %[[VAL_43]], %[[VAL_44]], %[[VAL_26]] : index 306// CHECK: scf.yield %[[VAL_42]], %[[VAL_45]], %[[VAL_46:.*]] : index, index, f32 307// CHECK: } 308// CHECK: memref.store %[[VAL_47:.*]]#2, %[[VAL_11]][] : memref<f32> 309// CHECK: %[[VAL_48:.*]] = bufferization.to_tensor %[[VAL_11]] : memref<f32> 310// CHECK: return %[[VAL_48]] : tensor<f32> 311// CHECK: } 312func.func @sparse_dot(%a: tensor<1024xf32, #SparseVector>, 313 %b: tensor<1024xf32, #SparseVector>, 314 %x: tensor<f32>) -> tensor<f32> { 315 %dot = linalg.dot ins(%a, %b: tensor<1024xf32, #SparseVector>, 316 tensor<1024xf32, #SparseVector>) 317 outs(%x: tensor<f32>) -> tensor<f32> 318 return %dot : tensor<f32> 319} 320