1 // RUN: mlir-opt %s --transform-interpreter --split-input-file | FileCheck %s
4 affine_map<(i, j, k) -> (i, k)>,
5 affine_map<(i, j, k) -> (k, j)>,
6 affine_map<(i, j, k) -> (i, j)>
9 indexing_maps = #matmat_accesses,
10 iterator_types = ["parallel", "parallel", "reduction"]
13 // CHECK-LABEL: func @matmul
14 // CHECK-SAME: %[[A:[a-zA-Z0-9]*]]: vector<2x4xf32>,
15 // CHECK-SAME: %[[B:[a-zA-Z0-9]*]]: vector<4x3xf32>,
16 // CHECK-SAME: %[[C:[a-zA-Z0-9]*]]: vector<2x3xf32>
17 // CHECK-DAG: %[[vcst:.*]] = arith.constant dense<0.000000e+00> : vector<8xf32>
18 // CHECK-DAG: %[[vcst_0:.*]] = arith.constant dense<0.000000e+00> : vector<12xf32>
19 // CHECK-DAG: %[[vcst_1:.*]] = arith.constant dense<0.000000e+00> : vector<2x3xf32>
20 // CHECK: %[[a0:.*]] = vector.extract %[[A]][0] : vector<4xf32> from vector<2x4xf32>
21 // CHECK: %[[a1:.*]] = vector.insert_strided_slice %[[a0]], %[[vcst]] {offsets = [0], strides = [1]} : vector<4xf32> into vector<8xf32>
22 // CHECK: %[[a2:.*]] = vector.extract %[[A]][1] : vector<4xf32> from vector<2x4xf32>
23 // CHECK: %[[a3:.*]] = vector.insert_strided_slice %[[a2]], %[[a1]] {offsets = [4], strides = [1]} : vector<4xf32> into vector<8xf32>
24 // CHECK: %[[b0:.*]] = vector.extract %[[B]][0] : vector<3xf32> from vector<4x3xf32>
25 // CHECK: %[[b1:.*]] = vector.insert_strided_slice %[[b0]], %[[vcst_0]] {offsets = [0], strides = [1]} : vector<3xf32> into vector<12xf32>
26 // CHECK: %[[b2:.*]] = vector.extract %[[B]][1] : vector<3xf32> from vector<4x3xf32>
27 // CHECK: %[[b3:.*]] = vector.insert_strided_slice %[[b2]], %[[b1]] {offsets = [3], strides = [1]} : vector<3xf32> into vector<12xf32>
28 // CHECK: %[[b4:.*]] = vector.extract %[[B]][2] : vector<3xf32> from vector<4x3xf32>
29 // CHECK: %[[b5:.*]] = vector.insert_strided_slice %[[b4]], %[[b3]] {offsets = [6], strides = [1]} : vector<3xf32> into vector<12xf32>
30 // CHECK: %[[b6:.*]] = vector.extract %[[B]][3] : vector<3xf32> from vector<4x3xf32>
31 // CHECK: %[[b7:.*]] = vector.insert_strided_slice %[[b6]], %[[b5]] {offsets = [9], strides = [1]} : vector<3xf32> into vector<12xf32>
32 // CHECK: %[[mm1:.*]] = vector.matrix_multiply %[[a3]], %[[b7]] {lhs_columns = 4 : i32, lhs_rows = 2 : i32, rhs_columns = 3 : i32} : (vector<8xf32>, vector<12xf32>) -> vector<6xf32>
33 // CHECK: %[[mm2:.*]] = vector.extract_strided_slice %[[mm1]] {offsets = [0], sizes = [3], strides = [1]} : vector<6xf32> to vector<3xf32>
34 // CHECK: %[[mm3:.*]] = vector.insert %[[mm2]], %[[vcst_1]] [0] : vector<3xf32> into vector<2x3xf32>
35 // CHECK: %[[mm4:.*]] = vector.extract_strided_slice %[[mm1]] {offsets = [3], sizes = [3], strides = [1]} : vector<6xf32> to vector<3xf32>
36 // CHECK: %[[mm5:.*]] = vector.insert %[[mm4]], %[[mm3]] [1] : vector<3xf32> into vector<2x3xf32>
37 // CHECK: %[[mm6:.*]] = arith.addf %[[C]], %[[mm5]] : vector<2x3xf32>
38 func.func @matmul(%arg0: vector<2x4xf32>,
39 %arg1: vector<4x3xf32>,
40 %arg2: vector<2x3xf32>) -> vector<2x3xf32> {
41 %0 = vector.contract #matmat_trait %arg0, %arg1, %arg2
42 : vector<2x4xf32>, vector<4x3xf32> into vector<2x3xf32>
43 return %0 : vector<2x3xf32>
46 // CHECK-LABEL: func @matmul_scalable
47 // CHECK-NOT: vector.matrix_multiply
48 func.func @matmul_scalable(%arg0: vector<2x4xf32>,
49 %arg1: vector<4x[3]xf32>,
50 %arg2: vector<2x[3]xf32>) -> vector<2x[3]xf32> {
51 %0 = vector.contract #matmat_trait %arg0, %arg1, %arg2
52 : vector<2x4xf32>, vector<4x[3]xf32> into vector<2x[3]xf32>
53 return %0 : vector<2x[3]xf32>
56 module attributes {transform.with_named_sequence} {
57 transform.named_sequence @__transform_main(%module_op: !transform.any_op {transform.readonly}) {
58 %f = transform.structured.match ops{["func.func"]} in %module_op
59 : (!transform.any_op) -> !transform.any_op
61 transform.apply_patterns to %f {
62 transform.apply_patterns.vector.lower_contraction lowering_strategy = "matmulintrinsics"
65 transform.apply_patterns to %f {
66 transform.apply_patterns.vector.lower_shape_cast