xref: /llvm-project/mlir/test/Integration/Dialect/Linalg/CPU/ArmSME/matmul.mlir (revision 8aeb104ce4d0907f2f1f5286611d6c2202d4ce53)
1// RUN: mlir-opt %s \
2// RUN:   -transform-interpreter -test-transform-dialect-erase-schedule \
3// RUN:   -test-lower-to-arm-sme -test-lower-to-llvm | \
4// RUN: %mcr_aarch64_cmd \
5// RUN:   -e=main -entry-point-result=void \
6// RUN:   -march=aarch64 -mattr="+sve,+sme" \
7// RUN:   -shared-libs=%native_mlir_runner_utils,%native_mlir_c_runner_utils,%native_arm_sme_abi_shlib | \
8// RUN: FileCheck %s
9
10func.func @matmul(%A : tensor<?x?xf32>, %B : tensor<?x?xf32>, %C : tensor<?x?xf32>) {
11  %res = linalg.matmul ins(%A, %B: tensor<?x?xf32>, tensor<?x?xf32>)
12                       outs(%C: tensor<?x?xf32>) -> tensor<?x?xf32>
13  %xf = tensor.cast %res : tensor<?x?xf32> to tensor<*xf32>
14  call @printMemrefF32(%xf) : (tensor<*xf32>) -> ()
15  return
16}
17
18func.func @main() {
19  %c0 = arith.constant 0 : i32
20  %c7 = arith.constant 7 : index
21
22  %A = arith.constant dense<[
23    [ 1.,  8., 15., 22., 29., 36., 43., 50., 57., 64., 71., 78., 85.],
24    [ 2.,  9., 16., 23., 30., 37., 44., 51., 58., 65., 72., 79., 86.],
25    [ 3., 10., 17., 24., 31., 38., 45., 52., 59., 66., 73., 80., 87.],
26    [ 4., 11., 18., 25., 32., 39., 46., 53., 60., 67., 74., 81., 88.],
27    [ 5., 12., 19., 26., 33., 40., 47., 54., 61., 68., 75., 82., 89.],
28    [ 6., 13., 20., 27., 34., 41., 48., 55., 62., 69., 76., 83., 90.],
29    [ 7., 14., 21., 28., 35., 42., 49., 56., 63., 70., 77., 84., 91.]
30  ]> : tensor<7x13xf32>
31
32  %B_init = tensor.empty() : tensor<13x7xf32>
33  %B = linalg.transpose ins(%A: tensor<7x13xf32>)
34                        outs(%B_init: tensor<13x7xf32>) permutation = [1, 0]
35
36  %A_dyn = tensor.cast %A : tensor<7x13xf32> to tensor<?x?xf32>
37  %B_dyn = tensor.cast %B : tensor<13x7xf32> to tensor<?x?xf32>
38
39  %C_init = bufferization.alloc_tensor(%c7, %c7) : tensor<?x?xf32>
40  %C = linalg.fill ins(%c0 : i32) outs(%C_init : tensor<?x?xf32>) -> tensor<?x?xf32>
41
42  // CHECK: Unranked Memref {{.*}} rank = 2 offset = 0 sizes = [7, 7] strides = [7, 1] data =
43  // CHECK: [32955, 33514, 34073, 34632, 35191, 35750, 36309]
44  // CHECK: [33514, 34086, 34658, 35230, 35802, 36374, 36946]
45  // CHECK: [34073, 34658, 35243, 35828, 36413, 36998, 37583]
46  // CHECK: [34632, 35230, 35828, 36426, 37024, 37622, 38220]
47  // CHECK: [35191, 35802, 36413, 37024, 37635, 38246, 38857]
48  // CHECK: [35750, 36374, 36998, 37622, 38246, 38870, 39494]
49  // CHECK: [36309, 36946, 37583, 38220, 38857, 39494, 40131]
50  call @matmul(%A_dyn, %B_dyn, %C) : (tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>) -> ()
51
52  return
53}
54
55module attributes {transform.with_named_sequence} {
56  transform.named_sequence @__transform_main(%module : !transform.any_op {transform.consumed}) {
57    %matmul = transform.structured.match ops{["linalg.matmul"]} in %module
58      : (!transform.any_op) -> !transform.any_op
59
60    // Step 1: Tile for size [4] x [4], which corresponds to SVLs x SVLs, where
61    // SVLs is the number of 32-bit elements in a vector of SVL bits.
62    %tiled_linalg_op, %loops:3 = transform.structured.tile_using_for %matmul tile_sizes [[4], [4], 1]
63      : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op, !transform.any_op)
64
65    // Step 2: Vectorize.
66    transform.structured.vectorize %tiled_linalg_op vector_sizes [[4], [4], 1]
67      : !transform.any_op
68
69    // Step 3: Bufferize ahead of TransferReadDropUnitDimsPattern, which
70    // currently only supports memrefs.
71    %bufferize = transform.bufferization.one_shot_bufferize %module
72      {bufferize_function_boundaries=true} : (!transform.any_op) -> !transform.any_op
73
74    %func = transform.structured.match ops{["func.func"]} in %bufferize
75      : (!transform.any_op) -> !transform.any_op
76
77    // Step 4: Lower vector.multi_reduction to vector.contract (+ some helpful patterns).
78    transform.apply_patterns to %func {
79      transform.apply_patterns.vector.lower_masked_transfers
80      transform.apply_patterns.vector.transfer_permutation_patterns
81      transform.apply_patterns.vector.reduction_to_contract
82    } : !transform.any_op
83
84    // Step 5: Lower vector.contract to vector.outerproduct. Also drop unit
85    // dims, specifically to prevent vector.transfer_read of vector<[4]x1xf32>,
86    // which can't be lowered in generic path.
87    transform.apply_patterns to %func {
88      transform.apply_patterns.vector.lower_contraction lowering_strategy = "outerproduct"
89      transform.apply_patterns.vector.lower_masks
90      transform.apply_patterns.vector.rank_reducing_subview_patterns
91      transform.apply_patterns.canonicalization
92    } : !transform.any_op
93
94    // Step 6 (optional optimization): Hoist accumulator load/store.
95    %func_h = transform.structured.hoist_redundant_vector_transfers %func
96        : (!transform.any_op) -> !transform.any_op
97    %all_loops = transform.structured.match interface{LoopLikeInterface} in %bufferize
98      : (!transform.any_op) -> !transform.any_op
99    transform.apply_licm to %all_loops : !transform.any_op
100    transform.loop.hoist_loop_invariant_subsets %all_loops : !transform.any_op
101    transform.yield
102  }
103}
104
105func.func private @printMemrefF32(%ptr : tensor<*xf32>)
106