1// RUN: mlir-opt %s \
2// RUN:   --sparsification --sparse-tensor-conversion \
3// RUN:   --convert-vector-to-scf --convert-scf-to-std \
4// RUN:   --func-bufferize --tensor-constant-bufferize --tensor-bufferize \
5// RUN:   --std-bufferize --finalizing-bufferize  \
6// RUN:   --convert-vector-to-llvm --convert-memref-to-llvm --convert-std-to-llvm --reconcile-unrealized-casts | \
7// RUN: TENSOR0="%mlir_integration_test_dir/data/test.mtx" \
8// RUN: mlir-cpu-runner \
9// RUN:  -e entry -entry-point-result=void  \
10// RUN:  -shared-libs=%mlir_integration_test_dir/libmlir_c_runner_utils%shlibext | \
11// RUN: FileCheck %s
12//
13// Do the same run, but now with SIMDization as well. This should not change the outcome.
14//
15// RUN: mlir-opt %s \
16// RUN:   --sparsification="vectorization-strategy=2 vl=4 enable-simd-index32" --sparse-tensor-conversion \
17// RUN:   --convert-vector-to-scf --convert-scf-to-std \
18// RUN:   --func-bufferize --tensor-constant-bufferize --tensor-bufferize \
19// RUN:   --std-bufferize --finalizing-bufferize --lower-affine \
20// RUN:   --convert-vector-to-llvm --convert-memref-to-llvm --convert-std-to-llvm --reconcile-unrealized-casts | \
21// RUN: TENSOR0="%mlir_integration_test_dir/data/test.mtx" \
22// RUN: mlir-cpu-runner \
23// RUN:  -e entry -entry-point-result=void  \
24// RUN:  -shared-libs=%mlir_integration_test_dir/libmlir_c_runner_utils%shlibext | \
25// RUN: FileCheck %s
26//
27
28!Filename = type !llvm.ptr<i8>
29
30#SparseMatrix = #sparse_tensor.encoding<{
31  dimLevelType = [ "compressed", "compressed" ],
32  pointerBitWidth = 32,
33  indexBitWidth = 32
34}>
35
36#trait_sampled_dense_dense = {
37  indexing_maps = [
38    affine_map<(i,j,k) -> (i,j)>,  // S
39    affine_map<(i,j,k) -> (i,k)>,  // A
40    affine_map<(i,j,k) -> (k,j)>,  // B
41    affine_map<(i,j,k) -> (i,j)>   // X (out)
42  ],
43  iterator_types = ["parallel", "parallel", "reduction"],
44  doc = "X(i,j) += S(i,j) SUM_k A(i,k) B(k,j)"
45}
46
47//
48// Integration test that lowers a kernel annotated as sparse to
49// actual sparse code, initializes a matching sparse storage scheme
50// from file, and runs the resulting code with the JIT compiler.
51//
52module {
53  //
54  // A kernel that computes a sampled matrix matrix multiplication.
55  //
56  func @sampled_dense_dense(%args: tensor<?x?xf32, #SparseMatrix>,
57                            %arga: tensor<?x?xf32>,
58                            %argb: tensor<?x?xf32>,
59                            %argx: tensor<?x?xf32> {linalg.inplaceable = true}) -> tensor<?x?xf32> {
60    %0 = linalg.generic #trait_sampled_dense_dense
61      ins(%args, %arga, %argb: tensor<?x?xf32, #SparseMatrix>, tensor<?x?xf32>, tensor<?x?xf32>)
62      outs(%argx: tensor<?x?xf32>) {
63        ^bb(%s: f32, %a: f32, %b: f32, %x: f32):
64          %0 = arith.mulf %a, %b : f32
65          %1 = arith.mulf %s, %0 : f32
66          %2 = arith.addf %x, %1 : f32
67          linalg.yield %2 : f32
68    } -> tensor<?x?xf32>
69    return %0 : tensor<?x?xf32>
70  }
71
72  func private @getTensorFilename(index) -> (!Filename)
73
74  //
75  // Main driver that reads matrix from file and calls the sparse kernel.
76  //
77  func @entry() {
78    %d0 = arith.constant 0.0 : f32
79    %c0 = arith.constant 0 : index
80    %c1 = arith.constant 1 : index
81    %c5 = arith.constant 5 : index
82    %c10 = arith.constant 10 : index
83
84    // Setup memory for the dense matrices and initialize.
85    %adata = memref.alloc(%c5, %c10) : memref<?x?xf32>
86    %bdata = memref.alloc(%c10, %c5) : memref<?x?xf32>
87    %xdata = memref.alloc(%c5,  %c5) : memref<?x?xf32>
88    scf.for %i = %c0 to %c5 step %c1 {
89      scf.for %j = %c0 to %c5 step %c1 {
90        memref.store %d0, %xdata[%i, %j] : memref<?x?xf32>
91      }
92      %p = arith.addi %i, %c1 : index
93      %q = arith.index_cast %p : index to i32
94      %d = arith.sitofp %q : i32 to f32
95      scf.for %j = %c0 to %c10 step %c1 {
96        memref.store %d, %adata[%i, %j] : memref<?x?xf32>
97        memref.store %d, %bdata[%j, %i] : memref<?x?xf32>
98      }
99    }
100    %a = bufferization.to_tensor %adata : memref<?x?xf32>
101    %b = bufferization.to_tensor %bdata : memref<?x?xf32>
102    %x = bufferization.to_tensor %xdata : memref<?x?xf32>
103
104    // Read the sparse matrix from file, construct sparse storage.
105    %fileName = call @getTensorFilename(%c0) : (index) -> (!Filename)
106    %s = sparse_tensor.new %fileName : !Filename to tensor<?x?xf32, #SparseMatrix>
107
108    // Call the kernel.
109    %0 = call @sampled_dense_dense(%s, %a, %b, %x)
110       : (tensor<?x?xf32, #SparseMatrix>,
111          tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>) -> tensor<?x?xf32>
112
113    // Print the result for verification.
114    //
115    // CHECK: ( 10, 0, 0, 56, 0 )
116    // CHECK: ( 0, 80, 0, 0, 250 )
117    // CHECK: ( 0, 0, 270, 0, 0 )
118    // CHECK: ( 164, 0, 0, 640, 0 )
119    // CHECK: ( 0, 520, 0, 0, 1250 )
120    //
121    %r = bufferization.to_memref %0 : memref<?x?xf32>
122    scf.for %i = %c0 to %c5 step %c1 {
123      %v = vector.transfer_read %r[%i, %c0], %d0: memref<?x?xf32>, vector<5xf32>
124      vector.print %v : vector<5xf32>
125    }
126
127    // Release the resources.
128    memref.dealloc %adata : memref<?x?xf32>
129    memref.dealloc %bdata : memref<?x?xf32>
130    memref.dealloc %xdata : memref<?x?xf32>
131    sparse_tensor.release %s : tensor<?x?xf32, #SparseMatrix>
132
133    return
134  }
135}
136