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#DCSR = #sparse_tensor.encoding<{ dimLevelType = [ "compressed", "compressed" ] }>
7
8// CHECK-LABEL:   func @matmul(
9// CHECK-SAME:                 %[[VAL_0:.*]]: tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>>,
10// CHECK-SAME:                 %[[VAL_1:.*]]: tensor<20x30xf32>,
11// CHECK-SAME:                 %[[VAL_2:.*]]: tensor<10x30xf32>) -> tensor<10x30xf32> {
12// CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 0 : index
13// CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 1 : index
14// CHECK-DAG:       %[[VAL_5:.*]] = arith.constant 30 : index
15// CHECK:           %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_0]], %[[VAL_3]] : tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>>
16// CHECK:           %[[VAL_7:.*]] = sparse_tensor.indices %[[VAL_0]], %[[VAL_3]] : tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>>
17// CHECK:           %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_0]], %[[VAL_4]] : tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>>
18// CHECK:           %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_0]], %[[VAL_4]] : tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>>
19// CHECK:           %[[VAL_10:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<10x20xf32, #sparse_tensor.encoding<{{{.*}}}>>
20// CHECK:           %[[VAL_11:.*]] = bufferization.to_memref %[[VAL_1]] : memref<20x30xf32>
21// CHECK:           %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_2]] : memref<10x30xf32>
22// CHECK:           %[[VAL_13:.*]] = memref.alloc() : memref<10x30xf32>
23// CHECK:           memref.copy %[[VAL_12]], %[[VAL_13]] : memref<10x30xf32> to 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 @matmul(%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// CHECK-LABEL:   func @conv2d(
56// CHECK-SAME:                 %[[VAL_0:.*]]: tensor<8x8xi32>,
57// CHECK-SAME:                 %[[VAL_1:.*]]: tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>>,
58// CHECK-SAME:                 %[[VAL_2:.*]]: tensor<6x6xi32>) -> tensor<6x6xi32> {
59// CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 0 : index
60// CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 1 : index
61// CHECK-DAG:       %[[VAL_5:.*]] = arith.constant 6 : index
62// CHECK:           %[[VAL_6:.*]] = bufferization.to_memref %[[VAL_0]] : memref<8x8xi32>
63// CHECK:           %[[VAL_7:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_3]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>>
64// CHECK:           %[[VAL_8:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_3]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>>
65// CHECK:           %[[VAL_9:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_4]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>>
66// CHECK:           %[[VAL_10:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_4]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>>
67// CHECK:           %[[VAL_11:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<3x3xi32, #sparse_tensor.encoding<{{{.*}}}>>
68// CHECK:           %[[VAL_12:.*]] = bufferization.to_memref %[[VAL_2]] : memref<6x6xi32>
69// CHECK:           %[[VAL_13:.*]] = memref.alloc() : memref<6x6xi32>
70// CHECK:           memref.copy %[[VAL_12]], %[[VAL_13]] : memref<6x6xi32> to memref<6x6xi32>
71// CHECK:           %[[VAL_14:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_3]]] : memref<?xindex>
72// CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_4]]] : memref<?xindex>
73// CHECK:           scf.for %[[VAL_16:.*]] = %[[VAL_14]] to %[[VAL_15]] step %[[VAL_4]] {
74// CHECK:             %[[VAL_17:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_16]]] : memref<?xindex>
75// CHECK:             %[[VAL_18:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_16]]] : memref<?xindex>
76// CHECK:             %[[VAL_19:.*]] = arith.addi %[[VAL_16]], %[[VAL_4]] : index
77// CHECK:             %[[VAL_20:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_19]]] : memref<?xindex>
78// CHECK:             scf.for %[[VAL_21:.*]] = %[[VAL_18]] to %[[VAL_20]] step %[[VAL_4]] {
79// CHECK:               %[[VAL_22:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_21]]] : memref<?xindex>
80// CHECK:               %[[VAL_23:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_21]]] : memref<?xi32>
81// CHECK:               scf.for %[[VAL_24:.*]] = %[[VAL_3]] to %[[VAL_5]] step %[[VAL_4]] {
82// CHECK:                 scf.for %[[VAL_25:.*]] = %[[VAL_3]] to %[[VAL_5]] step %[[VAL_4]] {
83// CHECK:                   %[[VAL_26:.*]] = memref.load %[[VAL_13]]{{\[}}%[[VAL_25]], %[[VAL_24]]] : memref<6x6xi32>
84// CHECK:                   %[[VAL_27:.*]] = arith.addi %[[VAL_25]], %[[VAL_17]] : index
85// CHECK:                   %[[VAL_28:.*]] = arith.addi %[[VAL_24]], %[[VAL_22]] : index
86// CHECK:                   %[[VAL_29:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_27]], %[[VAL_28]]] : memref<8x8xi32>
87// CHECK:                   %[[VAL_30:.*]] = arith.muli %[[VAL_29]], %[[VAL_23]] : i32
88// CHECK:                   %[[VAL_31:.*]] = arith.addi %[[VAL_26]], %[[VAL_30]] : i32
89// CHECK:                   memref.store %[[VAL_31]], %[[VAL_13]]{{\[}}%[[VAL_25]], %[[VAL_24]]] : memref<6x6xi32>
90// CHECK:                 }
91// CHECK:               }
92// CHECK:             }
93// CHECK:           }
94// CHECK:           %[[VAL_32:.*]] = bufferization.to_tensor %[[VAL_13]] : memref<6x6xi32>
95// CHECK:           return %[[VAL_32]] : tensor<6x6xi32>
96// CHECK:         }
97func @conv2d(%input:  tensor<8x8xi32>,
98             %filter: tensor<3x3xi32, #DCSR>,
99             %output: tensor<6x6xi32>) -> tensor<6x6xi32> {
100  %0 = linalg.conv_2d
101    ins  (%input, %filter: tensor<8x8xi32>, tensor<3x3xi32, #DCSR>)
102    outs (%output: tensor<6x6xi32>) -> tensor<6x6xi32>
103  return %0 : tensor<6x6xi32>
104}
105
106// CHECK-LABEL:   func @quantized_matmul(
107// CHECK-SAME:                           %[[VAL_0:.*]]: tensor<5x3xi8>,
108// CHECK-SAME:                           %[[VAL_1:.*]]: tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>>,
109// CHECK-SAME:                           %[[VAL_2:.*]]: tensor<5x6xi64>) -> tensor<5x6xi64> {
110// CHECK-DAG:       %[[VAL_3:.*]] = arith.constant 2 : i64
111// CHECK-DAG:       %[[VAL_4:.*]] = arith.constant 0 : index
112// CHECK-DAG:       %[[VAL_5:.*]] = arith.constant 1 : index
113// CHECK-DAG:       %[[VAL_6:.*]] = arith.constant 5 : index
114// CHECK:           %[[VAL_7:.*]] = bufferization.to_memref %[[VAL_0]] : memref<5x3xi8>
115// CHECK:           %[[VAL_8:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_4]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>>
116// CHECK:           %[[VAL_9:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_4]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>>
117// CHECK:           %[[VAL_10:.*]] = sparse_tensor.pointers %[[VAL_1]], %[[VAL_5]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>>
118// CHECK:           %[[VAL_11:.*]] = sparse_tensor.indices %[[VAL_1]], %[[VAL_5]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>>
119// CHECK:           %[[VAL_12:.*]] = sparse_tensor.values %[[VAL_1]] : tensor<3x6xi8, #sparse_tensor.encoding<{{{.*}}}>>
120// CHECK:           %[[VAL_13:.*]] = bufferization.to_memref %[[VAL_2]] : memref<5x6xi64>
121// CHECK:           %[[VAL_14:.*]] = memref.alloc() : memref<5x6xi64>
122// CHECK:           memref.copy %[[VAL_13]], %[[VAL_14]] : memref<5x6xi64> to memref<5x6xi64>
123// CHECK:           %[[VAL_15:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_4]]] : memref<?xindex>
124// CHECK:           %[[VAL_16:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_5]]] : memref<?xindex>
125// CHECK:           scf.for %[[VAL_17:.*]] = %[[VAL_15]] to %[[VAL_16]] step %[[VAL_5]] {
126// CHECK:             %[[VAL_18:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_17]]] : memref<?xindex>
127// CHECK:             %[[VAL_19:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_17]]] : memref<?xindex>
128// CHECK:             %[[VAL_20:.*]] = arith.addi %[[VAL_17]], %[[VAL_5]] : index
129// CHECK:             %[[VAL_21:.*]] = memref.load %[[VAL_10]]{{\[}}%[[VAL_20]]] : memref<?xindex>
130// CHECK:             scf.for %[[VAL_22:.*]] = %[[VAL_19]] to %[[VAL_21]] step %[[VAL_5]] {
131// CHECK:               %[[VAL_23:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_22]]] : memref<?xindex>
132// CHECK:               %[[VAL_24:.*]] = memref.load %[[VAL_12]]{{\[}}%[[VAL_22]]] : memref<?xi8>
133// CHECK:               scf.for %[[VAL_25:.*]] = %[[VAL_4]] to %[[VAL_6]] step %[[VAL_5]] {
134// CHECK:                 %[[VAL_26:.*]] = memref.load %[[VAL_14]]{{\[}}%[[VAL_25]], %[[VAL_23]]] : memref<5x6xi64>
135// CHECK:                 %[[VAL_27:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_25]], %[[VAL_18]]] : memref<5x3xi8>
136// CHECK:                 %[[VAL_28:.*]] = arith.extsi %[[VAL_27]] : i8 to i64
137// CHECK:                 %[[VAL_29:.*]] = arith.subi %[[VAL_28]], %[[VAL_3]] : i64
138// CHECK:                 %[[VAL_30:.*]] = arith.extsi %[[VAL_24]] : i8 to i64
139// CHECK:                 %[[VAL_31:.*]] = arith.muli %[[VAL_29]], %[[VAL_30]] : i64
140// CHECK:                 %[[VAL_32:.*]] = arith.addi %[[VAL_26]], %[[VAL_31]] : i64
141// CHECK:                 memref.store %[[VAL_32]], %[[VAL_14]]{{\[}}%[[VAL_25]], %[[VAL_23]]] : memref<5x6xi64>
142// CHECK:               }
143// CHECK:             }
144// CHECK:           }
145// CHECK:           %[[VAL_33:.*]] = bufferization.to_tensor %[[VAL_14]] : memref<5x6xi64>
146// CHECK:           return %[[VAL_33]] : tensor<5x6xi64>
147// CHECK:         }
148func @quantized_matmul(%input1: tensor<5x3xi8>,
149                       %input2: tensor<3x6xi8, #DCSR>,
150                       %output: tensor<5x6xi64>) -> tensor<5x6xi64> {
151  %c0 = arith.constant 0 : i32
152  %c2 = arith.constant 2 : i32
153  %0 = linalg.quantized_matmul
154    ins(%input1, %input2, %c2, %c0 : tensor<5x3xi8>, tensor<3x6xi8, #DCSR>, i32, i32)
155    outs(%output : tensor<5x6xi64>) -> tensor<5x6xi64>
156  return %0: tensor<5x6xi64>
157}
158