// RUN: mlir-opt %s \
// RUN: -transform-interpreter -test-transform-dialect-erase-schedule \
// RUN: -lower-vector-mask \
// RUN: -one-shot-bufferize="bufferize-function-boundaries" \
// RUN: -test-lower-to-arm-sme -test-lower-to-llvm | \
// RUN: %mcr_aarch64_cmd \
// RUN: -e=entry -entry-point-result=void \
// RUN: -march=aarch64 -mattr="+sve,+sme" \
// RUN: -shared-libs=%native_mlir_runner_utils,%native_mlir_c_runner_utils,%native_arm_sme_abi_shlib | \
// RUN: FileCheck %s
func.func @entry() {
%c0 = arith.constant 0 : index
%c4 = arith.constant 4 : index
%step = arith.constant 1 : index
%c123_f32 = arith.constant 123.0 : f32
// "svl" refers to the Streaming Vector Length and "svl_s" the number of
// 32-bit elements in a vector of SVL bits.
%svl_s = arm_sme.streaming_vl <word>
%tile_init = bufferization.alloc_tensor(%svl_s, %svl_s) : tensor<?x?xf32>
// Initialize tile with "123.0".
// TODO: this could be simplified to tensor.splat + tensor.insert_slice once
// splat supports dynamically shaped tensors.
%tile_0 = scf.for %i = %c0 to %svl_s step %step iter_args(%tile_partial = %tile_init) -> tensor<?x?xf32> {
%inner_tile = scf.for %j = %c0 to %svl_s step %step iter_args(%inner_tile_partial = %tile_partial) -> tensor<?x?xf32> {
%tile_update = tensor.insert %c123_f32 into %inner_tile_partial[%i, %j] : tensor<?x?xf32>
scf.yield %tile_update : tensor<?x?xf32>
}
scf.yield %inner_tile : tensor<?x?xf32>
}
// Print tile after initialization. The smallest SVL is 128-bits so the tile
// will be at least 4x4xf32.
//
// CHECK: ( 123, 123, 123, 123
// CHECK-NEXT: ( 123, 123, 123, 123
// CHECK-NEXT: ( 123, 123, 123, 123
// CHECK-NEXT: ( 123, 123, 123, 123
scf.for %i = %c0 to %svl_s step %step {
vector.print punctuation <open>
scf.for %j = %c0 to %svl_s step %step {
%element = tensor.extract %tile_0[%i, %j] : tensor<?x?xf32>
vector.print %element : f32 punctuation <no_punctuation>
// Print comma unless last element.
%c1_index = arith.constant 1 : index
%last_i = arith.subi %svl_s, %c1_index : index
%isNotLastIter = arith.cmpi ult, %j, %last_i : index
scf.if %isNotLastIter {
vector.print punctuation <comma>
}
}
vector.print punctuation <close>
vector.print punctuation <newline>
}
// Fill tile with pi.
%pi = arith.constant 3.14 : f32
%tile_1 = linalg.fill ins(%pi : f32) outs(%tile_0 : tensor<?x?xf32>) -> tensor<?x?xf32>
// Print tile after filling with pi. The smallest SVL is 128-bits so the tile
// will be at least 4x4xf32.
//
// CHECK: ( 3.14, 3.14, 3.14, 3.14
// CHECK-NEXT: ( 3.14, 3.14, 3.14, 3.14
// CHECK-NEXT: ( 3.14, 3.14, 3.14, 3.14
// CHECK-NEXT: ( 3.14, 3.14, 3.14, 3.14
scf.for %i = %c0 to %svl_s step %step {
vector.print punctuation <open>
scf.for %j = %c0 to %svl_s step %step {
%element = tensor.extract %tile_1[%i, %j] : tensor<?x?xf32>
vector.print %element : f32 punctuation <no_punctuation>
// Print comma unless last element.
%c1_index = arith.constant 1 : index
%last_i = arith.subi %svl_s, %c1_index : index
%isNotLastIter = arith.cmpi ult, %j, %last_i : index
scf.if %isNotLastIter {
vector.print punctuation <comma>
}
}
vector.print punctuation <close>
vector.print punctuation <newline>
}
// CHECK: SME: END OF TEST OUTPUT
vector.print str "SME: END OF TEST OUTPUT\n"
return
}
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["linalg.fill"]} in %arg1 : (!transform.any_op) -> !transform.any_op
transform.structured.vectorize %0 vector_sizes [[4], [4]] : !transform.any_op
transform.yield
}
}