xref: /llvm-project/mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_sampled_matmul.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// REDEFINE: %{env} = TENSOR0="%mlir_src_dir/test/Integration/data/test.mtx"
22// RUN: %{compile} | env %{env} %{run} | FileCheck %s
23//
24// Do the same run, but now with direct IR generation.
25// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false
26// RUN: %{compile} | env %{env} %{run} | FileCheck %s
27//
28// Do the same run, but now with direct IR generation and vectorization.
29// REDEFINE: %{sparsifier_opts} = enable-runtime-library=false vl=2 reassociate-fp-reductions=true enable-index-optimizations=true
30// RUN: %{compile} | env %{env} %{run} | FileCheck %s
31//
32// Do the same run, but now with direct IR generation and, if available, VLA
33// vectorization.
34// RUN: %if mlir_arm_sve_tests %{ %{compile_sve} | env %{env} %{run_sve} | FileCheck %s %}
35
36!Filename = !llvm.ptr
37
38#SparseMatrix = #sparse_tensor.encoding<{
39  map = (d0, d1) -> (d0 : compressed, d1 : compressed),
40  posWidth = 32,
41  crdWidth = 32
42}>
43
44#trait_sampled_dense_dense = {
45  indexing_maps = [
46    affine_map<(i,j,k) -> (i,j)>,  // S
47    affine_map<(i,j,k) -> (i,k)>,  // A
48    affine_map<(i,j,k) -> (k,j)>,  // B
49    affine_map<(i,j,k) -> (i,j)>   // X (out)
50  ],
51  iterator_types = ["parallel", "parallel", "reduction"],
52  doc = "X(i,j) += S(i,j) SUM_k A(i,k) B(k,j)"
53}
54
55//
56// Integration test that lowers a kernel annotated as sparse to
57// actual sparse code, initializes a matching sparse storage scheme
58// from file, and runs the resulting code with the JIT compiler.
59//
60module {
61  //
62  // A kernel that computes a sampled matrix matrix multiplication.
63  //
64  func.func @sampled_dense_dense(%args: tensor<?x?xf32, #SparseMatrix>,
65                                 %arga: tensor<?x?xf32>,
66                                 %argb: tensor<?x?xf32>,
67                                 %argx: tensor<?x?xf32>) -> tensor<?x?xf32> {
68    %0 = linalg.generic #trait_sampled_dense_dense
69      ins(%args, %arga, %argb: tensor<?x?xf32, #SparseMatrix>, tensor<?x?xf32>, tensor<?x?xf32>)
70      outs(%argx: tensor<?x?xf32>) {
71        ^bb(%s: f32, %a: f32, %b: f32, %x: f32):
72          %0 = arith.mulf %a, %b : f32
73          %1 = arith.mulf %s, %0 : f32
74          %2 = arith.addf %x, %1 : f32
75          linalg.yield %2 : f32
76    } -> tensor<?x?xf32>
77    return %0 : tensor<?x?xf32>
78  }
79
80  func.func private @getTensorFilename(index) -> (!Filename)
81
82  //
83  // Main driver that reads matrix from file and calls the sparse kernel.
84  //
85  func.func @main() {
86    %d0 = arith.constant 0.0 : f32
87    %c0 = arith.constant 0 : index
88    %c1 = arith.constant 1 : index
89    %c5 = arith.constant 5 : index
90    %c10 = arith.constant 10 : index
91
92    // Initialize dense matrices.
93    %x = tensor.generate %c5, %c5 {
94    ^bb0(%i : index, %j : index):
95      tensor.yield %d0 : f32
96    } : tensor<?x?xf32>
97
98    %a = tensor.generate %c5, %c10 {
99    ^bb0(%i: index, %j: index):
100      %p = arith.addi %i, %c1 : index
101      %q = arith.index_cast %p : index to i32
102      %d = arith.sitofp %q : i32 to f32
103      tensor.yield %d : f32
104    } : tensor<?x?xf32>
105
106    %b = tensor.generate %c10, %c5 {
107    ^bb0(%i: index, %j: index):
108      %p = arith.addi %j, %c1 : index
109      %q = arith.index_cast %p : index to i32
110      %d = arith.sitofp %q : i32 to f32
111      tensor.yield %d : f32
112    } : tensor<?x?xf32>
113
114    // Read the sparse matrix from file, construct sparse storage.
115    %fileName = call @getTensorFilename(%c0) : (index) -> (!Filename)
116    %s = sparse_tensor.new %fileName : !Filename to tensor<?x?xf32, #SparseMatrix>
117
118    // Call the kernel.
119    %0 = call @sampled_dense_dense(%s, %a, %b, %x)
120       : (tensor<?x?xf32, #SparseMatrix>,
121          tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>) -> tensor<?x?xf32>
122
123    // Print the result for verification.
124    //
125    // CHECK: ( 10, 0, 0, 56, 0 )
126    // CHECK: ( 0, 80, 0, 0, 250 )
127    // CHECK: ( 0, 0, 270, 0, 0 )
128    // CHECK: ( 164, 0, 0, 640, 0 )
129    // CHECK: ( 0, 520, 0, 0, 1250 )
130    //
131    scf.for %i = %c0 to %c5 step %c1 {
132      %v = vector.transfer_read %0[%i, %c0], %d0: tensor<?x?xf32>, vector<5xf32>
133      vector.print %v : vector<5xf32>
134    }
135
136    // Release the resources.
137    bufferization.dealloc_tensor %s : tensor<?x?xf32, #SparseMatrix>
138    bufferization.dealloc_tensor %0 : tensor<?x?xf32>
139    bufferization.dealloc_tensor %a : tensor<?x?xf32>
140    bufferization.dealloc_tensor %b : tensor<?x?xf32>
141
142    return
143  }
144}
145