xref: /llvm-project/mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_sampled_mm_fusion.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 direct IR generation and 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 direct IR generation and VLA vectorization.
32// RUN: %if mlir_arm_sve_tests %{ %{compile_sve} | %{run_sve} | FileCheck %s %}
33
34#SM = #sparse_tensor.encoding<{ map = (d0, d1) -> (d0 : compressed, d1 : compressed) }>
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#trait_matmul = {
48  indexing_maps = [
49    affine_map<(d0, d1, d2) -> (d1, d0)>,
50    affine_map<(d0, d1, d2) -> (d0, d2)>,
51    affine_map<(d0, d1, d2) -> (d1, d2)>
52  ],
53  iterator_types = ["reduction", "parallel", "parallel"]
54}
55
56#trait_scale = {
57  indexing_maps = [
58    affine_map<(d0, d1) -> (d0, d1)>,
59    affine_map<(d0, d1) -> (d0, d1)>,
60    affine_map<(d0, d1) -> (d0, d1)>
61  ],
62  iterator_types = ["parallel", "parallel"]
63}
64
65//
66// Integration test for sampled dense dense matmul fusion.
67//
68module {
69  //
70  // A kernel that computes a direct sampled matrix matrix multiplication
71  // (with dense result).
72  //
73  func.func @sampled_dd(%args: tensor<8x8xf64, #SM>,
74                        %arga: tensor<8x8xf64>,
75                        %argb: tensor<8x8xf64>) -> tensor<8x8xf64> {
76    %1 = arith.constant dense<0.0> : tensor<8x8xf64>
77    %2 = linalg.generic #trait_sampled_dense_dense
78      ins(%args, %arga, %argb: tensor<8x8xf64, #SM>,
79                               tensor<8x8xf64>, tensor<8x8xf64>)
80      outs(%1: tensor<8x8xf64>) {
81        ^bb(%s: f64, %a: f64, %b: f64, %x: f64):
82          %p = arith.mulf %a, %b : f64
83          %q = arith.mulf %s, %p : f64
84          %r = arith.addf %x, %q : f64
85          linalg.yield %r : f64
86    } -> tensor<8x8xf64>
87    return %2 : tensor<8x8xf64>
88  }
89
90  //
91  // A kernel that computes an unfused sampled matrix matrix multiplication
92  // (with dense result).
93  //
94  func.func @sampled_dd_unfused(%args: tensor<8x8xf64, #SM>,
95                                %arga: tensor<8x8xf64>,
96                                %argb: tensor<8x8xf64>) -> tensor<8x8xf64> {
97    // Perform dense-dense matrix matrix multiplication.
98    %1 = arith.constant dense<0.0> : tensor<8x8xf64>
99    %2 = linalg.generic #trait_matmul
100      ins(%arga, %argb : tensor<8x8xf64>, tensor<8x8xf64>)
101      outs(%1 : tensor<8x8xf64>) {
102        ^bb0(%a: f64, %b: f64, %x: f64):
103          %p = arith.mulf %a, %b : f64
104          %q = arith.addf %x, %p : f64
105          linalg.yield %q : f64
106    } -> tensor<8x8xf64>
107    // Sample the result with elements-wise multiplication with sparse matrix.
108    %3 = linalg.generic #trait_scale
109      ins(%2, %args : tensor<8x8xf64>, tensor<8x8xf64, #SM>)
110      outs(%1 : tensor<8x8xf64>) {
111        ^bb0(%t: f64, %s: f64, %x: f64):
112          %r = arith.mulf %t, %s : f64
113          linalg.yield %r : f64
114    } -> tensor<8x8xf64>
115    bufferization.dealloc_tensor %2 : tensor<8x8xf64>
116    return %3 : tensor<8x8xf64>
117  }
118
119  //
120  // A kernel that computes a direct sampled matrix matrix multiplication
121  // (with sparse result).
122  //
123  func.func @sparse_sampled_dd(%args: tensor<8x8xf64, #SM>,
124                               %arga: tensor<8x8xf64>,
125                               %argb: tensor<8x8xf64>) -> tensor<8x8xf64, #SM> {
126    %1 = tensor.empty() : tensor<8x8xf64, #SM>
127    %2 = linalg.generic #trait_sampled_dense_dense
128      ins(%args, %arga, %argb: tensor<8x8xf64, #SM>,
129                               tensor<8x8xf64>, tensor<8x8xf64>)
130      outs(%1: tensor<8x8xf64, #SM>) {
131        ^bb(%s: f64, %a: f64, %b: f64, %x: f64):
132          %p = arith.mulf %a, %b : f64
133          %q = arith.mulf %s, %p : f64
134          %r = arith.addf %x, %q : f64
135          linalg.yield %r : f64
136    } -> tensor<8x8xf64, #SM>
137    return %2 : tensor<8x8xf64, #SM>
138  }
139
140  //
141  // A kernel that computes an unfused sampled matrix matrix multiplication
142  // (with sparse result).
143  //
144  func.func @sparse_sampled_dd_unfused(
145        %args: tensor<8x8xf64, #SM>,
146        %arga: tensor<8x8xf64>,
147        %argb: tensor<8x8xf64>) -> tensor<8x8xf64, #SM> {
148    // Perform dense-dense matrix matrix multiplication.
149    %1 = arith.constant dense<0.0> : tensor<8x8xf64>
150    %2 = linalg.generic #trait_matmul
151      ins(%arga, %argb : tensor<8x8xf64>, tensor<8x8xf64>)
152      outs(%1 : tensor<8x8xf64>) {
153        ^bb0(%a: f64, %b: f64, %x: f64):
154          %p = arith.mulf %a, %b : f64
155          %q = arith.addf %x, %p : f64
156          linalg.yield %q : f64
157    } -> tensor<8x8xf64>
158    // Sample the result with elements-wise multiplication with sparse matrix.
159    %3 = tensor.empty() : tensor<8x8xf64, #SM>
160    %4 = linalg.generic #trait_scale
161      ins(%2, %args : tensor<8x8xf64>, tensor<8x8xf64, #SM>)
162      outs(%3 : tensor<8x8xf64, #SM>) {
163        ^bb0(%t: f64, %s: f64, %x: f64):
164          %r = arith.mulf %t, %s : f64
165          linalg.yield %r : f64
166    } -> tensor<8x8xf64, #SM>
167    return %4 : tensor<8x8xf64, #SM>
168  }
169
170  //
171  // Main driver.
172  //
173  func.func @main() {
174    %d0 = arith.constant 0.0 : f64
175    %c0 = arith.constant 0 : index
176
177    %t = arith.constant sparse<[[0, 0], [7,7]], [1.0, 2.0]>
178       : tensor<8x8xf64>
179    %s = sparse_tensor.convert %t
180       : tensor<8x8xf64> to tensor<8x8xf64, #SM>
181
182    %a = arith.constant dense<3.0> : tensor<8x8xf64>
183    %b = arith.constant dense<4.0> : tensor<8x8xf64>
184
185    // Call the kernels.
186    %0 = call @sampled_dd(%s, %a, %b)
187      : (tensor<8x8xf64, #SM>,
188         tensor<8x8xf64>, tensor<8x8xf64>) -> tensor<8x8xf64>
189    %1 = call @sampled_dd_unfused(%s, %a, %b)
190      : (tensor<8x8xf64, #SM>,
191         tensor<8x8xf64>, tensor<8x8xf64>) -> tensor<8x8xf64>
192    %2 = call @sparse_sampled_dd(%s, %a, %b)
193      : (tensor<8x8xf64, #SM>,
194         tensor<8x8xf64>, tensor<8x8xf64>) -> tensor<8x8xf64, #SM>
195    %3 = call @sparse_sampled_dd_unfused(%s, %a, %b)
196      : (tensor<8x8xf64, #SM>,
197         tensor<8x8xf64>, tensor<8x8xf64>) -> tensor<8x8xf64, #SM>
198
199    // Verify the outputs.
200    //
201    // CHECK:    ( ( 96, 0, 0, 0, 0, 0, 0, 0 ), ( 0, 0, 0, 0, 0, 0, 0, 0 ),
202    // CHECK-SAME: ( 0, 0, 0, 0, 0, 0, 0, 0 ), ( 0, 0, 0, 0, 0, 0, 0, 0 ),
203    // CHECK-SAME: ( 0, 0, 0, 0, 0, 0, 0, 0 ), ( 0, 0, 0, 0, 0, 0, 0, 0 ),
204    // CHECK-SAME: ( 0, 0, 0, 0, 0, 0, 0, 0 ), ( 0, 0, 0, 0, 0, 0, 0, 192 ) )
205    //
206    // CHECK:    ( ( 96, 0, 0, 0, 0, 0, 0, 0 ), ( 0, 0, 0, 0, 0, 0, 0, 0 ),
207    // CHECK-SAME: ( 0, 0, 0, 0, 0, 0, 0, 0 ), ( 0, 0, 0, 0, 0, 0, 0, 0 ),
208    // CHECK-SAME: ( 0, 0, 0, 0, 0, 0, 0, 0 ), ( 0, 0, 0, 0, 0, 0, 0, 0 ),
209    // CHECK-SAME: ( 0, 0, 0, 0, 0, 0, 0, 0 ), ( 0, 0, 0, 0, 0, 0, 0, 192 ) )
210    //
211    // CHECK:      ---- Sparse Tensor ----
212    // CHECK-NEXT: nse = 2
213    // CHECK-NEXT: dim = ( 8, 8 )
214    // CHECK-NEXT: lvl = ( 8, 8 )
215    // CHECK-NEXT: pos[0] : ( 0, 2 )
216    // CHECK-NEXT: crd[0] : ( 0, 7 )
217    // CHECK-NEXT: pos[1] : ( 0, 1, 2 )
218    // CHECK-NEXT: crd[1] : ( 0, 7 )
219    // CHECK-NEXT: values : ( 96, 192 )
220    // CHECK-NEXT: ----
221    //
222    // CHECK:      ---- Sparse Tensor ----
223    // CHECK-NEXT: nse = 2
224    // CHECK-NEXT: dim = ( 8, 8 )
225    // CHECK-NEXT: lvl = ( 8, 8 )
226    // CHECK-NEXT: pos[0] : ( 0, 2 )
227    // CHECK-NEXT: crd[0] : ( 0, 7 )
228    // CHECK-NEXT: pos[1] : ( 0, 1, 2 )
229    // CHECK-NEXT: crd[1] : ( 0, 7 )
230    // CHECK-NEXT: values : ( 96, 192 )
231    // CHECK-NEXT: ----
232    //
233    %v0 = vector.transfer_read %0[%c0, %c0], %d0
234        : tensor<8x8xf64>, vector<8x8xf64>
235    %v1 = vector.transfer_read %1[%c0, %c0], %d0
236        : tensor<8x8xf64>, vector<8x8xf64>
237    vector.print %v0 : vector<8x8xf64>
238    vector.print %v1 : vector<8x8xf64>
239    sparse_tensor.print %2 : tensor<8x8xf64, #SM>
240    sparse_tensor.print %3 : tensor<8x8xf64, #SM>
241
242    // Release the resources.
243    bufferization.dealloc_tensor %s : tensor<8x8xf64, #SM>
244    bufferization.dealloc_tensor %0 : tensor<8x8xf64>
245    bufferization.dealloc_tensor %1 : tensor<8x8xf64>
246    bufferization.dealloc_tensor %2 : tensor<8x8xf64, #SM>
247    bufferization.dealloc_tensor %3 : tensor<8x8xf64, #SM>
248
249    return
250  }
251}
252