Failure On Lower scf.parallel and scf.for

Beginners
1

Hi, I’m working on lowering the Linalg dialect to the LLVM Dialect. I’m trying to parallelize the matrix multiplication by OpenMP but have met some difficulties. My codes follow here.

func.func @matmul(%input: memref<2x2xf32>, %output: memref<2x2xf32>) {
  %init = arith.constant dense<[[1.0, 2.0], [3.0, 4.0]]>: tensor<2x2xf32>
  %init_buf = bufferization.to_memref %init: memref<2x2xf32>
  linalg.matmul ins(%input, %init_buf: memref<2x2xf32>, memref<2x2xf32>) outs(%output: memref<2x2xf32>)
  func.return
}

And my instruction follows.

mlir-opt-19 -convert-linalg-to-parallel-loops -convert-scf-to-openmp -convert-scf-to-cf matmul.mlir

It reports an error.

matmul.mlir:4:3: error: 'memref.alloca_scope' op expects region #0 to have 0 or 1 blocks
  linalg.matmul ins(%input, %init_buf: memref<2x2xf32>, memref<2x2xf32>) outs(%output: memref<2x2xf32>)
  ^
matmul.mlir:4:3: note: see current operation: 
"memref.alloca_scope"() ({
  "cf.br"(%2)[^bb1] : (index) -> ()
^bb1(%6: index):  // 2 preds: ^bb0, ^bb2
  %7 = "arith.cmpi"(%6, %1) <{predicate = 2 : i64}> : (index, index) -> i1
  "cf.cond_br"(%7)[^bb2, ^bb3] <{operandSegmentSizes = array<i32: 1, 0, 0>}> : (i1) -> ()
^bb2:  // pred: ^bb1
  %8 = "memref.load"(%arg0, %arg2, %6) <{nontemporal = false}> : (memref<2x2xf32>, index, index) -> f32
  %9 = "memref.load"(%4, %6, %arg3) <{nontemporal = false}> : (memref<2x2xf32>, index, index) -> f32
  %10 = "memref.load"(%arg1, %arg2, %arg3) <{nontemporal = false}> : (memref<2x2xf32>, index, index) -> f32
  %11 = "arith.mulf"(%8, %9) <{fastmath = #arith.fastmath<none>}> : (f32, f32) -> f32
  %12 = "arith.addf"(%10, %11) <{fastmath = #arith.fastmath<none>}> : (f32, f32) -> f32
  "memref.store"(%12, %arg1, %arg2, %arg3) <{nontemporal = false}> : (f32, memref<2x2xf32>, index, index) -> ()
  %13 = "arith.addi"(%6, %0) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
  "cf.br"(%13)[^bb1] : (index) -> ()
^bb3:  // pred: ^bb1
  "memref.alloca_scope.return"() : () -> ()
}) : () -> ()

I noticed it always happened when an “scf.for” is nested in the “scf.parallel”. Is there any way to make it work?

Thank you for your reading and patience!

2

Going through the conversation in Help lowering affine loop to OpenMP might help the memref.alloca_scope issue.

3

Thank you for your reply. Your link is beneficial. I made some modifications to it, and it worked.

mlir-opt-19 -lower-affine -convert-scf-to-openmp -convert-func-to-llvm -arith-bufferize -finalize-memref-to-llvm -convert-scf-to-cf -convert-openmp-to-llvm -canonicalize -convert-to-llvm matmul.mlir
1 Like