Is there a way to randomize memref or assign values to memref one by one?

MLIR
1

The memory allocated by memref on spike simulator is all zeros.I want to do some arithmetic and run some tests. Is there a way to initialize memref? Because I’m allocating a lot of memory.If anyone could help me, I would be grateful.Thanks!

2

Hi, mlir::linalg::FillRng2DOp from the linalg dialect should help you with that. For constant values use the mlir::linalg::FillOp.

Alternatively, you could initialize it manually.

%alloc = memref.alloc() : memref<4x4xf64>
%cst = arith.constant 0.000000e+00 : f64 // or some other value
affine.for %arg0 = 0 to 4 {
  affine.for %arg1 = 0 to 4 {
    affine.store %cst, %alloc[%arg0, %arg1] : memref<4x4xf64>
  }
}
2 Likes
3

Bear in mind that FillRng2DOp lowers code in the IR to do the RNG.

@philipportner’s manual initialization is more or less what FillOp would do.

Another option is to use dense globals:

memref.global "private" constant @__constant : memref<2x2x3x8xf32> = dense<1.000000e+00> {alignment = 128 : i64}
...
%1 = memref.get_global @__constant : memref<2x2x3x8xf32>

At tensor level, you can just declare an arith.constant (which will become a global+get after bufferization):

%cst = arith.constant dense<4.900000e+01> : tensor<2x2x3x8xf32>

The dense attribute doesn’t need to be a splat (all of the same value), but if not, you have to add every single value there, which can be messy. There’s also a hex representation that compacts things a bit if you have large tensors.

2 Likes