xref: /llvm-project/mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_pack.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#SortedCOO = #sparse_tensor.encoding<{
35  map = (d0, d1) -> (d0 : compressed(nonunique), d1 : singleton)
36}>
37
38#SortedCOOI32 = #sparse_tensor.encoding<{
39  map = (d0, d1) -> (d0 : compressed(nonunique), d1 : singleton),
40  posWidth = 32,
41  crdWidth = 32
42}>
43
44#CSR = #sparse_tensor.encoding<{
45  map = (d0, d1) -> (d0 : dense, d1 : compressed),
46  posWidth = 32,
47  crdWidth = 32
48}>
49
50#BCOO = #sparse_tensor.encoding<{
51  map = (d0, d1, d2) -> (d0 : dense, d1 : loose_compressed(nonunique), d2 : singleton)
52}>
53
54module {
55  //
56  // Main driver.
57  //
58  func.func @main() {
59    %c0 = arith.constant 0 : index
60    %f0 = arith.constant 0.0 : f64
61    %i0 = arith.constant 0 : i32
62
63    //
64    // Setup COO.
65    //
66
67    %data = arith.constant dense<
68       [ 1.0,  2.0,  3.0 ]
69    > : tensor<3xf64>
70
71    %pos = arith.constant dense<
72       [0, 3]
73    > : tensor<2xindex>
74
75    %index = arith.constant dense<
76       [[  1,  2],
77        [  5,  6],
78        [  7,  8]]
79    > : tensor<3x2xindex>
80
81    %pos32 = arith.constant dense<
82       [0, 3]
83    > : tensor<2xi32>
84
85    %index32 = arith.constant dense<
86       [[  1,  2],
87        [  5,  6],
88        [  7,  8]]
89    > : tensor<3x2xi32>
90
91    %s4 = sparse_tensor.assemble (%pos, %index), %data : (tensor<2xindex>, tensor<3x2xindex>), tensor<3xf64>
92                                          to tensor<10x10xf64, #SortedCOO>
93    %s5 = sparse_tensor.assemble (%pos32, %index32), %data : (tensor<2xi32>, tensor<3x2xi32>), tensor<3xf64>
94                                          to tensor<10x10xf64, #SortedCOOI32>
95
96    //
97    // Setup CSR.
98    //
99
100    %csr_data = arith.constant dense<
101       [ 1.0,  2.0,  3.0 ]
102    > : tensor<3xf64>
103
104    %csr_pos32 = arith.constant dense<
105       [0, 1, 3]
106    > : tensor<3xi32>
107
108    %csr_index32 = arith.constant dense<
109       [1, 0, 1]
110    > : tensor<3xi32>
111    %csr = sparse_tensor.assemble (%csr_pos32, %csr_index32), %csr_data : (tensor<3xi32>, tensor<3xi32>), tensor<3xf64>
112                                           to tensor<2x2xf64, #CSR>
113
114    //
115    // Setup BCOO.
116    //
117
118    %bdata = arith.constant dense<
119       [ 1.0,  2.0,  3.0,  4.0,  5.0 ]
120    > : tensor<5xf64>
121
122    %bpos = arith.constant dense<
123       [0, 3, 3, 5]
124    > : tensor<4xindex>
125
126    %bindex = arith.constant dense<
127      [[  1,  2],
128       [  5,  6],
129       [  7,  8],
130       [  2,  3],
131       [  4,  2],
132       [ 10, 10]]
133    > : tensor<6x2xindex>
134
135    %bs = sparse_tensor.assemble (%bpos, %bindex), %bdata :
136          (tensor<4xindex>,  tensor<6x2xindex>), tensor<5xf64> to tensor<2x10x10xf64, #BCOO>
137
138    //
139    // Verify results.
140    //
141
142    // CHECK:     1
143    // CHECK-NEXT:2
144    // CHECK-NEXT:1
145    //
146    // CHECK-NEXT:5
147    // CHECK-NEXT:6
148    // CHECK-NEXT:2
149    //
150    // CHECK-NEXT:7
151    // CHECK-NEXT:8
152    // CHECK-NEXT:3
153    sparse_tensor.foreach in %s4 : tensor<10x10xf64, #SortedCOO> do {
154      ^bb0(%1: index, %2: index, %v: f64) :
155        vector.print %1: index
156        vector.print %2: index
157        vector.print %v: f64
158    }
159
160    // CHECK-NEXT:1
161    // CHECK-NEXT:2
162    // CHECK-NEXT:1
163    //
164    // CHECK-NEXT:5
165    // CHECK-NEXT:6
166    // CHECK-NEXT:2
167    //
168    // CHECK-NEXT:7
169    // CHECK-NEXT:8
170    // CHECK-NEXT:3
171    sparse_tensor.foreach in %s5 : tensor<10x10xf64, #SortedCOOI32> do {
172      ^bb0(%1: index, %2: index, %v: f64) :
173        vector.print %1: index
174        vector.print %2: index
175        vector.print %v: f64
176    }
177
178    // CHECK-NEXT:0
179    // CHECK-NEXT:1
180    // CHECK-NEXT:1
181    //
182    // CHECK-NEXT:1
183    // CHECK-NEXT:0
184    // CHECK-NEXT:2
185    //
186    // CHECK-NEXT:1
187    // CHECK-NEXT:1
188    // CHECK-NEXT:3
189    sparse_tensor.foreach in %csr : tensor<2x2xf64, #CSR> do {
190      ^bb0(%1: index, %2: index, %v: f64) :
191        vector.print %1: index
192        vector.print %2: index
193        vector.print %v: f64
194    }
195
196    // CHECK-NEXT:0
197    // CHECK-NEXT:1
198    // CHECK-NEXT:2
199    // CHECK-NEXT:1
200    //
201    // CHECK-NEXT:0
202    // CHECK-NEXT:5
203    // CHECK-NEXT:6
204    // CHECK-NEXT:2
205    //
206    // CHECK-NEXT:0
207    // CHECK-NEXT:7
208    // CHECK-NEXT:8
209    // CHECK-NEXT:3
210    //
211    // CHECK-NEXT:1
212    // CHECK-NEXT:2
213    // CHECK-NEXT:3
214    // CHECK-NEXT:4
215    //
216    // CHECK-NEXT:1
217    // CHECK-NEXT:4
218    // CHECK-NEXT:2
219    // CHECK-NEXT:5
220    sparse_tensor.foreach in %bs : tensor<2x10x10xf64, #BCOO> do {
221      ^bb0(%0: index, %1: index, %2: index, %v: f64) :
222        vector.print %0: index
223        vector.print %1: index
224        vector.print %2: index
225        vector.print %v: f64
226    }
227
228    //
229    // Verify disassemble operations.
230    //
231
232    %od = tensor.empty() : tensor<3xf64>
233    %op = tensor.empty() : tensor<2xi32>
234    %oi = tensor.empty() : tensor<3x2xi32>
235    %p, %i, %d, %pl, %il, %dl = sparse_tensor.disassemble %s5 : tensor<10x10xf64, #SortedCOOI32>
236                 out_lvls(%op, %oi : tensor<2xi32>, tensor<3x2xi32>)
237                 out_vals(%od : tensor<3xf64>)
238                 -> (tensor<2xi32>, tensor<3x2xi32>), tensor<3xf64>, (i32, i64), index
239
240    // CHECK-NEXT: ( 1, 2, 3 )
241    %vd = vector.transfer_read %d[%c0], %f0 : tensor<3xf64>, vector<3xf64>
242    vector.print %vd : vector<3xf64>
243
244    // CHECK-NEXT: ( ( 1, 2 ), ( 5, 6 ), ( 7, 8 ) )
245    %vi = vector.transfer_read %i[%c0, %c0], %i0 : tensor<3x2xi32>, vector<3x2xi32>
246    vector.print %vi : vector<3x2xi32>
247
248    // CHECK-NEXT: 3
249    vector.print %dl : index
250
251    %d_csr = tensor.empty() : tensor<4xf64>
252    %p_csr = tensor.empty() : tensor<3xi32>
253    %i_csr = tensor.empty() : tensor<3xi32>
254    %rp_csr, %ri_csr, %rd_csr, %lp_csr, %li_csr, %ld_csr = sparse_tensor.disassemble %csr : tensor<2x2xf64, #CSR>
255                 out_lvls(%p_csr, %i_csr : tensor<3xi32>, tensor<3xi32>)
256                 out_vals(%d_csr : tensor<4xf64>)
257                 -> (tensor<3xi32>, tensor<3xi32>), tensor<4xf64>, (i32, i64), index
258
259    // CHECK-NEXT: ( 1, 2, 3 )
260    %vd_csr = vector.transfer_read %rd_csr[%c0], %f0 : tensor<4xf64>, vector<3xf64>
261    vector.print %vd_csr : vector<3xf64>
262
263    // CHECK-NEXT: 3
264    vector.print %ld_csr : index
265
266    %bod = tensor.empty() : tensor<6xf64>
267    %bop = tensor.empty() : tensor<4xindex>
268    %boi = tensor.empty() : tensor<6x2xindex>
269    %bp, %bi, %bd, %lp, %li, %ld = sparse_tensor.disassemble %bs : tensor<2x10x10xf64, #BCOO>
270                    out_lvls(%bop, %boi : tensor<4xindex>, tensor<6x2xindex>)
271                    out_vals(%bod : tensor<6xf64>)
272                    -> (tensor<4xindex>, tensor<6x2xindex>), tensor<6xf64>, (i32, tensor<i64>), index
273
274    // CHECK-NEXT: ( 1, 2, 3, 4, 5 )
275    %vbd = vector.transfer_read %bd[%c0], %f0 : tensor<6xf64>, vector<5xf64>
276    vector.print %vbd : vector<5xf64>
277
278    // CHECK-NEXT: 5
279    vector.print %ld : index
280
281    // CHECK-NEXT: ( ( 1, 2 ), ( 5, 6 ), ( 7, 8 ), ( 2, 3 ), ( 4, 2 ), ( {{.*}}, {{.*}} ) )
282    %vbi = vector.transfer_read %bi[%c0, %c0], %c0 : tensor<6x2xindex>, vector<6x2xindex>
283    vector.print %vbi : vector<6x2xindex>
284
285    // CHECK-NEXT: 10
286    %si = tensor.extract %li[] : tensor<i64>
287    vector.print %si : i64
288
289    // TODO: This check is no longer needed once the codegen path uses the
290    // buffer deallocation pass. "dealloc_tensor" turn into a no-op in the
291    // codegen path.
292    %has_runtime = sparse_tensor.has_runtime_library
293    scf.if %has_runtime {
294      // sparse_tensor.assemble copies buffers when running with the runtime
295      // library. Deallocations are not needed when running in codegen mode.
296      bufferization.dealloc_tensor %s4 : tensor<10x10xf64, #SortedCOO>
297      bufferization.dealloc_tensor %s5 : tensor<10x10xf64, #SortedCOOI32>
298      bufferization.dealloc_tensor %csr : tensor<2x2xf64, #CSR>
299      bufferization.dealloc_tensor %bs : tensor<2x10x10xf64, #BCOO>
300    }
301
302    bufferization.dealloc_tensor %li : tensor<i64>
303    bufferization.dealloc_tensor %od : tensor<3xf64>
304    bufferization.dealloc_tensor %op : tensor<2xi32>
305    bufferization.dealloc_tensor %oi : tensor<3x2xi32>
306    bufferization.dealloc_tensor %d_csr : tensor<4xf64>
307    bufferization.dealloc_tensor %p_csr : tensor<3xi32>
308    bufferization.dealloc_tensor %i_csr : tensor<3xi32>
309    bufferization.dealloc_tensor %bod : tensor<6xf64>
310    bufferization.dealloc_tensor %bop : tensor<4xindex>
311    bufferization.dealloc_tensor %boi : tensor<6x2xindex>
312
313    return
314  }
315}
316