1// RUN: mlir-opt %s -sparsification | FileCheck %s --check-prefix=CHECK-HIR
2//
3// RUN: mlir-opt %s -sparsification --sparse-tensor-conversion | \
4// RUN: FileCheck %s --check-prefix=CHECK-MIR
5//
6// RUN: mlir-opt %s -sparsification --sparse-tensor-conversion \
7// RUN: --func-bufferize --tensor-constant-bufferize           \
8// RUN: --tensor-bufferize --finalizing-bufferize |            \
9// RUN: FileCheck %s --check-prefix=CHECK-LIR
10
11#CSR = #sparse_tensor.encoding<{dimLevelType = [ "dense", "compressed" ]}>
12
13#trait_matvec = {
14  indexing_maps = [
15    affine_map<(i,j) -> (i,j)>,  // A
16    affine_map<(i,j) -> (j)>,    // b
17    affine_map<(i,j) -> (i)>     // x (out)
18  ],
19  iterator_types = ["parallel","reduction"],
20  doc = "x(i) += A(i,j) * b(j)"
21}
22
23// CHECK-HIR-LABEL:   func @matvec(
24// CHECK-HIR-SAME:                 %[[VAL_0:.*]]: tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ], pointerBitWidth = 0, indexBitWidth = 0 }>>,
25// CHECK-HIR-SAME:                 %[[VAL_1:.*]]: tensor<64xf64>,
26// CHECK-HIR-SAME:                 %[[VAL_2:.*]]: tensor<32xf64>) -> tensor<32xf64> {
27// CHECK-HIR:           %[[VAL_3:.*]] = constant 32 : index
28// CHECK-HIR:           %[[VAL_4:.*]] = constant 0 : index
29// CHECK-HIR:           %[[VAL_5:.*]] = constant 1 : index
30// CHECK-HIR:           %[[VAL_6:.*]] = sparse_tensor.pointers %[[VAL_0]], %[[VAL_5]] : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ], pointerBitWidth = 0, indexBitWidth = 0 }>> to memref<?xindex>
31// CHECK-HIR:           %[[VAL_7:.*]] = sparse_tensor.indices %[[VAL_0]], %[[VAL_5]] : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ], pointerBitWidth = 0, indexBitWidth = 0 }>> to memref<?xindex>
32// CHECK-HIR:           %[[VAL_8:.*]] = sparse_tensor.values %[[VAL_0]] : tensor<32x64xf64, #sparse_tensor.encoding<{ dimLevelType = [ "dense", "compressed" ], pointerBitWidth = 0, indexBitWidth = 0 }>> to memref<?xf64>
33// CHECK-HIR:           %[[VAL_9:.*]] = memref.buffer_cast %[[VAL_1]] : memref<64xf64>
34// CHECK-HIR:           %[[VAL_10:.*]] = memref.buffer_cast %[[VAL_2]] : memref<32xf64>
35// CHECK-HIR:           %[[VAL_11:.*]] = memref.alloc() : memref<32xf64>
36// CHECK-HIR:           memref.copy %[[VAL_10]], %[[VAL_11]] : memref<32xf64> to memref<32xf64>
37// CHECK-HIR:           scf.for %[[VAL_12:.*]] = %[[VAL_4]] to %[[VAL_3]] step %[[VAL_5]] {
38// CHECK-HIR:             %[[VAL_13:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_12]]] : memref<?xindex>
39// CHECK-HIR:             %[[VAL_14:.*]] = addi %[[VAL_12]], %[[VAL_5]] : index
40// CHECK-HIR:             %[[VAL_15:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_14]]] : memref<?xindex>
41// CHECK-HIR:             %[[VAL_16:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_12]]] : memref<32xf64>
42// CHECK-HIR:             %[[VAL_17:.*]] = scf.for %[[VAL_18:.*]] = %[[VAL_13]] to %[[VAL_15]] step %[[VAL_5]] iter_args(%[[VAL_19:.*]] = %[[VAL_16]]) -> (f64) {
43// CHECK-HIR:               %[[VAL_20:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_18]]] : memref<?xindex>
44// CHECK-HIR:               %[[VAL_21:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_18]]] : memref<?xf64>
45// CHECK-HIR:               %[[VAL_22:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_20]]] : memref<64xf64>
46// CHECK-HIR:               %[[VAL_23:.*]] = mulf %[[VAL_21]], %[[VAL_22]] : f64
47// CHECK-HIR:               %[[VAL_24:.*]] = addf %[[VAL_19]], %[[VAL_23]] : f64
48// CHECK-HIR:               scf.yield %[[VAL_24]] : f64
49// CHECK-HIR:             }
50// CHECK-HIR:             memref.store %[[VAL_25:.*]], %[[VAL_11]]{{\[}}%[[VAL_12]]] : memref<32xf64>
51// CHECK-HIR:           }
52// CHECK-HIR:           %[[VAL_26:.*]] = memref.tensor_load %[[VAL_11]] : memref<32xf64>
53// CHECK-HIR:           return %[[VAL_26]] : tensor<32xf64>
54// CHECK-HIR:         }
55
56// CHECK-MIR-LABEL:   func @matvec(
57// CHECK-MIR-SAME:                 %[[VAL_0:.*]]: !llvm.ptr<i8>,
58// CHECK-MIR-SAME:                 %[[VAL_1:.*]]: tensor<64xf64>,
59// CHECK-MIR-SAME:                 %[[VAL_2:.*]]: tensor<32xf64>) -> tensor<32xf64> {
60// CHECK-MIR:           %[[VAL_3:.*]] = constant 32 : index
61// CHECK-MIR:           %[[VAL_4:.*]] = constant 0 : index
62// CHECK-MIR:           %[[VAL_5:.*]] = constant 1 : index
63// CHECK-MIR:           %[[VAL_6:.*]] = call @sparsePointers(%[[VAL_0]], %[[VAL_5]]) : (!llvm.ptr<i8>, index) -> memref<?xindex>
64// CHECK-MIR:           %[[VAL_7:.*]] = call @sparseIndices(%[[VAL_0]], %[[VAL_5]]) : (!llvm.ptr<i8>, index) -> memref<?xindex>
65// CHECK-MIR:           %[[VAL_8:.*]] = call @sparseValuesF64(%[[VAL_0]]) : (!llvm.ptr<i8>) -> memref<?xf64>
66// CHECK-MIR:           %[[VAL_9:.*]] = memref.buffer_cast %[[VAL_1]] : memref<64xf64>
67// CHECK-MIR:           %[[VAL_10:.*]] = memref.buffer_cast %[[VAL_2]] : memref<32xf64>
68// CHECK-MIR:           %[[VAL_11:.*]] = memref.alloc() : memref<32xf64>
69// CHECK-MIR:           memref.copy %[[VAL_10]], %[[VAL_11]] : memref<32xf64> to memref<32xf64>
70// CHECK-MIR:           scf.for %[[VAL_14:.*]] = %[[VAL_4]] to %[[VAL_3]] step %[[VAL_5]] {
71// CHECK-MIR:             %[[VAL_15:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_14]]] : memref<?xindex>
72// CHECK-MIR:             %[[VAL_16:.*]] = addi %[[VAL_14]], %[[VAL_5]] : index
73// CHECK-MIR:             %[[VAL_17:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_16]]] : memref<?xindex>
74// CHECK-MIR:             %[[VAL_18:.*]] = memref.load %[[VAL_11]]{{\[}}%[[VAL_14]]] : memref<32xf64>
75// CHECK-MIR:             %[[VAL_19:.*]] = scf.for %[[VAL_20:.*]] = %[[VAL_15]] to %[[VAL_17]] step %[[VAL_5]] iter_args(%[[VAL_21:.*]] = %[[VAL_18]]) -> (f64) {
76// CHECK-MIR:               %[[VAL_22:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_20]]] : memref<?xindex>
77// CHECK-MIR:               %[[VAL_23:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_20]]] : memref<?xf64>
78// CHECK-MIR:               %[[VAL_24:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_22]]] : memref<64xf64>
79// CHECK-MIR:               %[[VAL_25:.*]] = mulf %[[VAL_23]], %[[VAL_24]] : f64
80// CHECK-MIR:               %[[VAL_26:.*]] = addf %[[VAL_21]], %[[VAL_25]] : f64
81// CHECK-MIR:               scf.yield %[[VAL_26]] : f64
82// CHECK-MIR:             }
83// CHECK-MIR:             memref.store %[[VAL_27:.*]], %[[VAL_11]]{{\[}}%[[VAL_14]]] : memref<32xf64>
84// CHECK-MIR:           }
85// CHECK-MIR:           %[[VAL_28:.*]] = memref.tensor_load %[[VAL_11]] : memref<32xf64>
86// CHECK-MIR:           return %[[VAL_28]] : tensor<32xf64>
87// CHECK-MIR:         }
88
89// CHECK-LIR-LABEL:   func @matvec(
90// CHECK-LIR-SAME:                 %[[VAL_0:.*]]: !llvm.ptr<i8>,
91// CHECK-LIR-SAME:                 %[[VAL_1:.*]]: memref<64xf64>,
92// CHECK-LIR-SAME:                 %[[VAL_2:.*]]: memref<32xf64>) -> memref<32xf64> {
93// CHECK-LIR:           %[[VAL_3:.*]] = constant 32 : index
94// CHECK-LIR:           %[[VAL_4:.*]] = constant 0 : index
95// CHECK-LIR:           %[[VAL_5:.*]] = constant 1 : index
96// CHECK-LIR:           %[[VAL_6:.*]] = call @sparsePointers(%[[VAL_0]], %[[VAL_5]]) : (!llvm.ptr<i8>, index) -> memref<?xindex>
97// CHECK-LIR:           %[[VAL_7:.*]] = call @sparseIndices(%[[VAL_0]], %[[VAL_5]]) : (!llvm.ptr<i8>, index) -> memref<?xindex>
98// CHECK-LIR:           %[[VAL_8:.*]] = call @sparseValuesF64(%[[VAL_0]]) : (!llvm.ptr<i8>) -> memref<?xf64>
99// CHECK-LIR:           %[[VAL_9:.*]] = memref.alloc() : memref<32xf64>
100// CHECK-LIR:           memref.copy %[[VAL_2]], %[[VAL_9]] : memref<32xf64> to memref<32xf64>
101// CHECK-LIR:           scf.for %[[VAL_12:.*]] = %[[VAL_4]] to %[[VAL_3]] step %[[VAL_5]] {
102// CHECK-LIR:             %[[VAL_13:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_12]]] : memref<?xindex>
103// CHECK-LIR:             %[[VAL_14:.*]] = addi %[[VAL_12]], %[[VAL_5]] : index
104// CHECK-LIR:             %[[VAL_15:.*]] = memref.load %[[VAL_6]]{{\[}}%[[VAL_14]]] : memref<?xindex>
105// CHECK-LIR:             %[[VAL_16:.*]] = memref.load %[[VAL_9]]{{\[}}%[[VAL_12]]] : memref<32xf64>
106// CHECK-LIR:             %[[VAL_17:.*]] = scf.for %[[VAL_18:.*]] = %[[VAL_13]] to %[[VAL_15]] step %[[VAL_5]] iter_args(%[[VAL_19:.*]] = %[[VAL_16]]) -> (f64) {
107// CHECK-LIR:               %[[VAL_20:.*]] = memref.load %[[VAL_7]]{{\[}}%[[VAL_18]]] : memref<?xindex>
108// CHECK-LIR:               %[[VAL_21:.*]] = memref.load %[[VAL_8]]{{\[}}%[[VAL_18]]] : memref<?xf64>
109// CHECK-LIR:               %[[VAL_22:.*]] = memref.load %[[VAL_1]]{{\[}}%[[VAL_20]]] : memref<64xf64>
110// CHECK-LIR:               %[[VAL_23:.*]] = mulf %[[VAL_21]], %[[VAL_22]] : f64
111// CHECK-LIR:               %[[VAL_24:.*]] = addf %[[VAL_19]], %[[VAL_23]] : f64
112// CHECK-LIR:               scf.yield %[[VAL_24]] : f64
113// CHECK-LIR:             }
114// CHECK-LIR:             memref.store %[[VAL_25:.*]], %[[VAL_9]]{{\[}}%[[VAL_12]]] : memref<32xf64>
115// CHECK-LIR:           }
116// CHECK-LIR:           return %[[VAL_9]] : memref<32xf64>
117// CHECK-LIR:         }
118
119func @matvec(%arga: tensor<32x64xf64, #CSR>,
120             %argb: tensor<64xf64>,
121             %argx: tensor<32xf64>) -> tensor<32xf64> {
122  %0 = linalg.generic #trait_matvec
123      ins(%arga, %argb : tensor<32x64xf64, #CSR>, tensor<64xf64>)
124      outs(%argx: tensor<32xf64>) {
125    ^bb(%A: f64, %b: f64, %x: f64):
126      %0 = mulf %A, %b : f64
127      %1 = addf %x, %0 : f64
128      linalg.yield %1 : f64
129  } -> tensor<32xf64>
130  return %0 : tensor<32xf64>
131}
132