MLIR for multiplying 2 matrices - llvm

Question

I am trying to multiply 2 matrices using MLIR. However, IR does not compile and gives this error:
./matmult.mlir:24:16: error: custom op 'linalg.fill' [parseNamedStructuredOpRegion] ods-gen generated region expects 2 args, got 0
The error happens in this line (linalg.fill(%A, %cf1) : memref<2048x2048xf64>, f64). Despite two args are provided to linalg.fill, it gives error of no args. I am using LLVM16 and this command to compile (mlir-opt -convert-linalg-to-loops ./matmult.mlir)
MLIR file:
// C += A * B.
func.func #matmul(%A: memref<2048x2048xf64>, %B: memref<2048x2048xf64>, %C: memref<2048x2048xf64>) {
affine.for %arg3 = 0 to 2048 {
affine.for %arg4 = 0 to 2048 {
affine.for %arg5 = 0 to 2048 {
%a = affine.load %A[%arg3, %arg5] : memref<2048x2048xf64>
%b = affine.load %B[%arg5, %arg4] : memref<2048x2048xf64>
%ci = affine.load %C[%arg3, %arg4] : memref<2048x2048xf64>
%p = arith.mulf %a, %b : f64
%co = arith.addf %ci, %p : f64
affine.store %co, %C[%arg3, %arg4] : memref<2048x2048xf64>
}
}
}
return
}
func.func #main() {
%A = memref.alloc() : memref<2048x2048xf64>
%B = memref.alloc() : memref<2048x2048xf64>
%C = memref.alloc() : memref<2048x2048xf64>
%cf1 = llvm.mlir.constant(1.00000e+00 : f64) : f64
linalg.fill(%A, %cf1) : memref<2048x2048xf64>, f64
linalg.fill(%B, %cf1) : memref<2048x2048xf64>, f64
linalg.fill(%C, %cf1) : memref<2048x2048xf64>, f64
call #matmul(%A, %B, %C) : (memref<2048x2048xf64>, memref<2048x2048xf64>, memref<2048x2048xf64>) -> ()
call #print_memref_2d_f64(%C): (memref<2048x2048xf64>) -> ()
return
}
func.func #print_memref_2d_f64(memref<2048x2048xf64>)

Answer 1

Instruction notation seems to be changed, replacing all of those linalg.fill with,
linalg.fill ins(%cf1 : f64) outs(%A : memref<2048x2048xf64>)
linalg.fill ins(%cf1 : f64) outs(%B : memref<2048x2048xf64>)
linalg.fill ins(%cf1 : f64) outs(%C : memref<2048x2048xf64>)
Produces:
module {
func.func #matmul(%arg0: memref<2048x2048xf64>, %arg1: memref<2048x2048xf64>, %arg2: memref<2048x2048xf64>) {
affine.for %arg3 = 0 to 2048 {
affine.for %arg4 = 0 to 2048 {
affine.for %arg5 = 0 to 2048 {
%0 = affine.load %arg0[%arg3, %arg5] : memref<2048x2048xf64>
%1 = affine.load %arg1[%arg5, %arg4] : memref<2048x2048xf64>
%2 = affine.load %arg2[%arg3, %arg4] : memref<2048x2048xf64>
%3 = arith.mulf %0, %1 : f64
%4 = arith.addf %2, %3 : f64
affine.store %4, %arg2[%arg3, %arg4] : memref<2048x2048xf64>
}
}
}
return
}
func.func #main() {
%c2048 = arith.constant 2048 : index
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%0 = llvm.mlir.constant(1.000000e+00 : f64) : f64
%1 = memref.alloc() : memref<2048x2048xf64>
%2 = memref.alloc() : memref<2048x2048xf64>
%3 = memref.alloc() : memref<2048x2048xf64>
scf.for %arg0 = %c0 to %c2048 step %c1 {
scf.for %arg1 = %c0 to %c2048 step %c1 {
memref.store %0, %1[%arg0, %arg1] : memref<2048x2048xf64>
}
}
scf.for %arg0 = %c0 to %c2048 step %c1 {
scf.for %arg1 = %c0 to %c2048 step %c1 {
memref.store %0, %2[%arg0, %arg1] : memref<2048x2048xf64>
}
}
scf.for %arg0 = %c0 to %c2048 step %c1 {
scf.for %arg1 = %c0 to %c2048 step %c1 {
memref.store %0, %3[%arg0, %arg1] : memref<2048x2048xf64>
}
}
call #matmul(%1, %2, %3) : (memref<2048x2048xf64>, memref<2048x2048xf64>, memref<2048x2048xf64>) -> ()
return
}
}
(I commented out call to print_memref_2d_f64 function as it's missing)
I took usage syntax from here