xref: /llvm-project/mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_index_dense.mlir (revision eb206e9ea84eff0a0596fed2de8316d924f946d1)
1//--------------------------------------------------------------------------------------------------
2// WHEN CREATING A NEW TEST, PLEASE JUST COPY & PASTE WITHOUT EDITS.
3//
4// Set-up that's shared across all tests in this directory. In principle, this
5// config could be moved to lit.local.cfg. However, there are downstream users that
6//  do not use these LIT config files. Hence why this is kept inline.
7//
8// DEFINE: %{sparsifier_opts} = enable-runtime-library=true
9// DEFINE: %{sparsifier_opts_sve} = enable-arm-sve=true %{sparsifier_opts}
10// DEFINE: %{compile} = mlir-opt %s --sparsifier="%{sparsifier_opts}"
11// DEFINE: %{compile_sve} = mlir-opt %s --sparsifier="%{sparsifier_opts_sve}"
12// DEFINE: %{run_libs} = -shared-libs=%mlir_c_runner_utils,%mlir_runner_utils
13// DEFINE: %{run_libs_sve} = -shared-libs=%native_mlir_runner_utils,%native_mlir_c_runner_utils
14// DEFINE: %{run_opts} = -e main -entry-point-result=void
15// DEFINE: %{run} = mlir-runner %{run_opts} %{run_libs}
16// DEFINE: %{run_sve} = %mcr_aarch64_cmd --march=aarch64 --mattr="+sve" %{run_opts} %{run_libs_sve}
17//
18// DEFINE: %{env} =
19//--------------------------------------------------------------------------------------------------
20
21// RUN: %{compile} | %{run} | FileCheck %s
22//
23// Do the same run, but now with direct IR generation.
24// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false
25// RUN: %{compile} | %{run} | FileCheck %s
26//
27// Do the same run, but now with direct IR generation and vectorization.
28// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false vl=2 reassociate-fp-reductions=true enable-index-optimizations=true
29// RUN: %{compile} | %{run} | FileCheck %s
30//
31// Do the same run, but now with direct IR generation and VLA vectorization.
32// RUN: %if mlir_arm_sve_tests %{ %{compile_sve} | %{run_sve} | FileCheck %s %}
33
34#SparseVector = #sparse_tensor.encoding<{
35  map = (d0) -> (d0 : compressed)
36}>
37
38#SparseMatrix = #sparse_tensor.encoding<{
39  map = (d0, d1) -> (d0 : compressed, d1 : compressed)
40}>
41
42#trait_1d = {
43  indexing_maps = [
44    affine_map<(i) -> (i)>,  // a
45    affine_map<(i) -> (i)>   // x (out)
46  ],
47  iterator_types = ["parallel"],
48  doc = "X(i) = a(i) op i"
49}
50
51#trait_2d = {
52  indexing_maps = [
53    affine_map<(i,j) -> (i,j)>,  // A
54    affine_map<(i,j) -> (i,j)>   // X (out)
55  ],
56  iterator_types = ["parallel", "parallel"],
57  doc = "X(i,j) = A(i,j) op i op j"
58}
59
60//
61// Test with indices and sparse inputs. All outputs are dense.
62//
63module {
64
65  //
66  // Kernel that uses index in the index notation (conjunction).
67  //
68  func.func @sparse_index_1d_conj(%arga: tensor<8xi64, #SparseVector>,
69                                  %out: tensor<8xi64>) -> tensor<8xi64> {
70    %r = linalg.generic #trait_1d
71        ins(%arga: tensor<8xi64, #SparseVector>)
72       outs(%out: tensor<8xi64>) {
73        ^bb(%a: i64, %x: i64):
74          %i = linalg.index 0 : index
75          %ii = arith.index_cast %i : index to i64
76          %m1 = arith.muli %a, %ii : i64
77          linalg.yield %m1 : i64
78    } -> tensor<8xi64>
79    return %r : tensor<8xi64>
80  }
81
82  //
83  // Kernel that uses index in the index notation (disjunction).
84  //
85  func.func @sparse_index_1d_disj(%arga: tensor<8xi64, #SparseVector>,
86                                  %out: tensor<8xi64>) -> tensor<8xi64> {
87    %r = linalg.generic #trait_1d
88        ins(%arga: tensor<8xi64, #SparseVector>)
89       outs(%out: tensor<8xi64>) {
90        ^bb(%a: i64, %x: i64):
91          %i = linalg.index 0 : index
92          %ii = arith.index_cast %i : index to i64
93          %m1 = arith.addi %a, %ii : i64
94          linalg.yield %m1 : i64
95    } -> tensor<8xi64>
96    return %r : tensor<8xi64>
97  }
98
99  //
100  // Kernel that uses indices in the index notation (conjunction).
101  //
102  func.func @sparse_index_2d_conj(%arga: tensor<3x4xi64, #SparseMatrix>,
103                                  %out: tensor<3x4xi64>) -> tensor<3x4xi64> {
104    %r = linalg.generic #trait_2d
105        ins(%arga: tensor<3x4xi64, #SparseMatrix>)
106       outs(%out: tensor<3x4xi64>) {
107        ^bb(%a: i64, %x: i64):
108          %i = linalg.index 0 : index
109          %j = linalg.index 1 : index
110          %ii = arith.index_cast %i : index to i64
111          %jj = arith.index_cast %j : index to i64
112          %m1 = arith.muli %ii, %a : i64
113          %m2 = arith.muli %jj, %m1 : i64
114          linalg.yield %m2 : i64
115    } -> tensor<3x4xi64>
116    return %r : tensor<3x4xi64>
117  }
118
119  //
120  // Kernel that uses indices in the index notation (disjunction).
121  //
122  func.func @sparse_index_2d_disj(%arga: tensor<3x4xi64, #SparseMatrix>,
123                                  %out: tensor<3x4xi64>) -> tensor<3x4xi64> {
124    %r = linalg.generic #trait_2d
125        ins(%arga: tensor<3x4xi64, #SparseMatrix>)
126       outs(%out: tensor<3x4xi64>) {
127        ^bb(%a: i64, %x: i64):
128          %i = linalg.index 0 : index
129          %j = linalg.index 1 : index
130          %ii = arith.index_cast %i : index to i64
131          %jj = arith.index_cast %j : index to i64
132          %m1 = arith.addi %ii, %a : i64
133          %m2 = arith.addi %jj, %m1 : i64
134          linalg.yield %m2 : i64
135    } -> tensor<3x4xi64>
136    return %r : tensor<3x4xi64>
137  }
138
139  //
140  // Main driver.
141  //
142  func.func @main() {
143    %c0 = arith.constant 0 : index
144    %du = arith.constant -1 : i64
145
146    // Setup input sparse vector.
147    %v1 = arith.constant sparse<[[2], [4]], [ 10, 20]> : tensor<8xi64>
148    %sv = sparse_tensor.convert %v1 : tensor<8xi64> to tensor<8xi64, #SparseVector>
149
150    // Setup input "sparse" vector.
151    %v2 = arith.constant dense<[ 1,  2,  4,  8,  16,  32,  64,  128 ]> : tensor<8xi64>
152    %dv = sparse_tensor.convert %v2 : tensor<8xi64> to tensor<8xi64, #SparseVector>
153
154    // Setup input sparse matrix.
155    %m1 = arith.constant sparse<[[1,1], [2,3]], [10, 20]> : tensor<3x4xi64>
156    %sm = sparse_tensor.convert %m1 : tensor<3x4xi64> to tensor<3x4xi64, #SparseMatrix>
157
158    // Setup input "sparse" matrix.
159    %m2 = arith.constant dense <[ [ 1,  1,  1,  1 ],
160                                  [ 1,  2,  1,  1 ],
161                                  [ 1,  1,  3,  4 ] ]> : tensor<3x4xi64>
162    %dm = sparse_tensor.convert %m2 : tensor<3x4xi64> to tensor<3x4xi64, #SparseMatrix>
163
164    // Setup out tensors.
165    // Note: Constants bufferize to read-only buffers.
166    %init_8 = tensor.empty() : tensor<8xi64>
167    %init_3_4 = tensor.empty() : tensor<3x4xi64>
168
169    // Call the kernels.
170    %0 = call @sparse_index_1d_conj(%sv, %init_8) : (tensor<8xi64, #SparseVector>, tensor<8xi64>) -> tensor<8xi64>
171    %1 = call @sparse_index_1d_disj(%sv, %init_8) : (tensor<8xi64, #SparseVector>, tensor<8xi64>) -> tensor<8xi64>
172    %2 = call @sparse_index_1d_conj(%dv, %init_8) : (tensor<8xi64, #SparseVector>, tensor<8xi64>) -> tensor<8xi64>
173    %3 = call @sparse_index_1d_disj(%dv, %init_8) : (tensor<8xi64, #SparseVector>, tensor<8xi64>) -> tensor<8xi64>
174    %4 = call @sparse_index_2d_conj(%sm, %init_3_4) : (tensor<3x4xi64, #SparseMatrix>, tensor<3x4xi64>) -> tensor<3x4xi64>
175    %5 = call @sparse_index_2d_disj(%sm, %init_3_4) : (tensor<3x4xi64, #SparseMatrix>, tensor<3x4xi64>) -> tensor<3x4xi64>
176    %6 = call @sparse_index_2d_conj(%dm, %init_3_4) : (tensor<3x4xi64, #SparseMatrix>, tensor<3x4xi64>) -> tensor<3x4xi64>
177    %7 = call @sparse_index_2d_disj(%dm, %init_3_4) : (tensor<3x4xi64, #SparseMatrix>, tensor<3x4xi64>) -> tensor<3x4xi64>
178
179    //
180    // Verify result.
181    //
182    // CHECK:      ( 0, 0, 20, 0, 80, 0, 0, 0 )
183    // CHECK-NEXT: ( 0, 1, 12, 3, 24, 5, 6, 7 )
184    // CHECK-NEXT: ( 0, 2, 8, 24, 64, 160, 384, 896 )
185    // CHECK-NEXT: ( 1, 3, 6, 11, 20, 37, 70, 135 )
186    // CHECK-NEXT: ( ( 0, 0, 0, 0 ), ( 0, 10, 0, 0 ), ( 0, 0, 0, 120 ) )
187    // CHECK-NEXT: ( ( 0, 1, 2, 3 ), ( 1, 12, 3, 4 ), ( 2, 3, 4, 25 ) )
188    // CHECK-NEXT: ( ( 0, 0, 0, 0 ), ( 0, 2, 2, 3 ), ( 0, 2, 12, 24 ) )
189    // CHECK-NEXT: ( ( 1, 2, 3, 4 ), ( 2, 4, 4, 5 ), ( 3, 4, 7, 9 ) )
190    //
191    %vv0 = vector.transfer_read %0[%c0], %du: tensor<8xi64>, vector<8xi64>
192    %vv1 = vector.transfer_read %1[%c0], %du: tensor<8xi64>, vector<8xi64>
193    %vv2 = vector.transfer_read %2[%c0], %du: tensor<8xi64>, vector<8xi64>
194    %vv3 = vector.transfer_read %3[%c0], %du: tensor<8xi64>, vector<8xi64>
195    %vv4 = vector.transfer_read %4[%c0,%c0], %du: tensor<3x4xi64>, vector<3x4xi64>
196    %vv5 = vector.transfer_read %5[%c0,%c0], %du: tensor<3x4xi64>, vector<3x4xi64>
197    %vv6 = vector.transfer_read %6[%c0,%c0], %du: tensor<3x4xi64>, vector<3x4xi64>
198    %vv7 = vector.transfer_read %7[%c0,%c0], %du: tensor<3x4xi64>, vector<3x4xi64>
199    vector.print %vv0 : vector<8xi64>
200    vector.print %vv1 : vector<8xi64>
201    vector.print %vv2 : vector<8xi64>
202    vector.print %vv3 : vector<8xi64>
203    vector.print %vv4 : vector<3x4xi64>
204    vector.print %vv5 : vector<3x4xi64>
205    vector.print %vv6 : vector<3x4xi64>
206    vector.print %vv7 : vector<3x4xi64>
207
208    // Release resources.
209    bufferization.dealloc_tensor %sv : tensor<8xi64, #SparseVector>
210    bufferization.dealloc_tensor %dv : tensor<8xi64, #SparseVector>
211    bufferization.dealloc_tensor %sm : tensor<3x4xi64, #SparseMatrix>
212    bufferization.dealloc_tensor %dm : tensor<3x4xi64, #SparseMatrix>
213    bufferization.dealloc_tensor %0 : tensor<8xi64>
214    bufferization.dealloc_tensor %1 : tensor<8xi64>
215    bufferization.dealloc_tensor %2 : tensor<8xi64>
216    bufferization.dealloc_tensor %3 : tensor<8xi64>
217    bufferization.dealloc_tensor %4 : tensor<3x4xi64>
218    bufferization.dealloc_tensor %5 : tensor<3x4xi64>
219    bufferization.dealloc_tensor %6 : tensor<3x4xi64>
220    bufferization.dealloc_tensor %7 : tensor<3x4xi64>
221
222    return
223  }
224}
225