xref: /llvm-project/mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_transpose.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 enable-buffer-initialization=true
25// RUN: %{compile} | %{run} | FileCheck %s
26//
27// Do the same run, but now with vectorization.
28// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false enable-buffer-initialization=true 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  VLA vectorization.
32// RUN: %if mlir_arm_sve_tests %{ %{compile_sve} | %{run_sve} | FileCheck %s %}
33
34#DCSR = #sparse_tensor.encoding<{
35  map = (d0, d1) -> (d0 : compressed, d1 : compressed)
36}>
37
38#DCSC = #sparse_tensor.encoding<{
39  map = (d0, d1) -> (d1 : compressed, d0 : compressed)
40}>
41
42#transpose_trait = {
43  indexing_maps = [
44    affine_map<(i,j) -> (j,i)>,  // A
45    affine_map<(i,j) -> (i,j)>   // X
46  ],
47  iterator_types = ["parallel", "parallel"],
48  doc = "X(i,j) = A(j,i)"
49}
50
51module {
52
53  //
54  // Transposing a sparse row-wise matrix into another sparse row-wise
55  // matrix introduces a cycle in the iteration graph. This complication
56  // can be avoided by manually inserting a conversion of the incoming
57  // matrix into a sparse column-wise matrix first.
58  //
59  func.func @sparse_transpose(%arga: tensor<3x4xf64, #DCSR>)
60                                  -> tensor<4x3xf64, #DCSR> {
61    %t = sparse_tensor.convert %arga
62      : tensor<3x4xf64, #DCSR> to tensor<3x4xf64, #DCSC>
63
64    %i = tensor.empty() : tensor<4x3xf64, #DCSR>
65    %0 = linalg.generic #transpose_trait
66       ins(%t: tensor<3x4xf64, #DCSC>)
67       outs(%i: tensor<4x3xf64, #DCSR>) {
68       ^bb(%a: f64, %x: f64):
69         linalg.yield %a : f64
70    } -> tensor<4x3xf64, #DCSR>
71
72    bufferization.dealloc_tensor %t : tensor<3x4xf64, #DCSC>
73
74    return %0 : tensor<4x3xf64, #DCSR>
75  }
76
77  //
78  // However, even better, the sparsifier is able to insert such a
79  // conversion automatically to resolve a cycle in the iteration graph!
80  //
81  func.func @sparse_transpose_auto(%arga: tensor<3x4xf64, #DCSR>)
82                                       -> tensor<4x3xf64, #DCSR> {
83    %i = tensor.empty() : tensor<4x3xf64, #DCSR>
84    %0 = linalg.generic #transpose_trait
85       ins(%arga: tensor<3x4xf64, #DCSR>)
86       outs(%i: tensor<4x3xf64, #DCSR>) {
87       ^bb(%a: f64, %x: f64):
88         linalg.yield %a : f64
89    } -> tensor<4x3xf64, #DCSR>
90    return %0 : tensor<4x3xf64, #DCSR>
91  }
92
93  //
94  // Main driver.
95  //
96  func.func @main() {
97    %c0 = arith.constant 0 : index
98    %c1 = arith.constant 1 : index
99    %c4 = arith.constant 4 : index
100    %du = arith.constant 0.0 : f64
101
102    // Setup input sparse matrix from compressed constant.
103    %d = arith.constant dense <[
104       [ 1.1,  1.2,  0.0,  1.4 ],
105       [ 0.0,  0.0,  0.0,  0.0 ],
106       [ 3.1,  0.0,  3.3,  3.4 ]
107    ]> : tensor<3x4xf64>
108    %a = sparse_tensor.convert %d : tensor<3x4xf64> to tensor<3x4xf64, #DCSR>
109
110    // Call the kernels.
111    %0 = call @sparse_transpose(%a)
112      : (tensor<3x4xf64, #DCSR>) -> tensor<4x3xf64, #DCSR>
113    %1 = call @sparse_transpose_auto(%a)
114      : (tensor<3x4xf64, #DCSR>) -> tensor<4x3xf64, #DCSR>
115
116    //
117    // Verify result.
118    //
119    // CHECK:      ---- Sparse Tensor ----
120    // CHECK-NEXT: nse = 6
121    // CHECK-NEXT: dim = ( 4, 3 )
122    // CHECK-NEXT: lvl = ( 4, 3 )
123    // CHECK-NEXT: pos[0] : ( 0, 4 )
124    // CHECK-NEXT: crd[0] : ( 0, 1, 2, 3 )
125    // CHECK-NEXT: pos[1] : ( 0, 2, 3, 4, 6 )
126    // CHECK-NEXT: crd[1] : ( 0, 2, 0, 2, 0, 2 )
127    // CHECK-NEXT: values : ( 1.1, 3.1, 1.2, 3.3, 1.4, 3.4 )
128    // CHECK-NEXT: ----
129    // CHECK:      ---- Sparse Tensor ----
130    // CHECK-NEXT: nse = 6
131    // CHECK-NEXT: dim = ( 4, 3 )
132    // CHECK-NEXT: lvl = ( 4, 3 )
133    // CHECK-NEXT: pos[0] : ( 0, 4 )
134    // CHECK-NEXT: crd[0] : ( 0, 1, 2, 3 )
135    // CHECK-NEXT: pos[1] : ( 0, 2, 3, 4, 6 )
136    // CHECK-NEXT: crd[1] : ( 0, 2, 0, 2, 0, 2 )
137    // CHECK-NEXT: values : ( 1.1, 3.1, 1.2, 3.3, 1.4, 3.4 )
138    // CHECK-NEXT: ----
139    //
140    sparse_tensor.print %0 : tensor<4x3xf64, #DCSR>
141    sparse_tensor.print %1 : tensor<4x3xf64, #DCSR>
142
143    // Release resources.
144    bufferization.dealloc_tensor %a : tensor<3x4xf64, #DCSR>
145    bufferization.dealloc_tensor %0 : tensor<4x3xf64, #DCSR>
146    bufferization.dealloc_tensor %1 : tensor<4x3xf64, #DCSR>
147
148    return
149  }
150}
151