Need help for bufferization

MLIR
1

sample IR works fine.

func.func @sample() {
 %1 = arith.constant dense<[3]>  : tensor<1xi8>
 return
}

 memref.global "private" constant @__constant_1xi8 : memref<1xi8> = dense<3> {alignment = 64 : i64}
 func.func @sample() {
   %0 = memref.get_global @__constant_1xi8 : memref<1xi8>
   return
 }

But is there elegant way to specify the space of memref for constant op ?
defaultMemorySpaceFn works, unfortunately, it can NOT specify different space for different constant ops.

Tensor encoding is alternavite way, but it doesn’t work for constant op :frowning:

func.func @sample() {
 %1 = arith.constant dense<[3]>  : tensor<1xi8, 2>
 return
}

error: 'memref.global' op initial value expected to be of type 'tensor<1xi8>', but was of type 'tensor<1xi8, 2 : i64>'

does anyone know other methods?
thank you very much.

2

Tensor IR specifies what to compute, but not where the data is located. If you want a buffer in a specific memory space, I’d recommend to put an explicit allocation. E.g.:

memref.global "private" constant @__constant_1xi8 : memref<1xi8, 2> = dense<3> {alignment = 64 : i64}
 func.func @sample() {
   %0 = memref.get_global @__constant_1xi8 : memref<1xi8, 2>
   %1 = bufferization.to_tensor %0 restrict : memref<1xi8, 2>
   return
 }

Make sure to put the restrict attribute. Otherwise, One-Shot Bufferize will reject the IR.

1 Like
3

yes, i understand the difference regarding tensor and buffer(memref).
the issue is about how to inject memory space info. for this buffer in the process of bufferization.

the method i can think of is based on encoding information of Tensor Type

// input MLIR
func.func @sample() {
 %1 = arith.constant dense<[3]>  : tensor<1xi8, 2>  // 2 is expected memory space
 return
}

// command
mlir-opt --one-shot-bufferize="use-encoding-for-memory-space" input.mlir -o output.mlir

// Expected output MLIR
 memref.global "private" constant @__constant_1xi8 : memref<1xi8, 2> = dense<3> {alignment = 64 : i64}
 func.func @sample() {
   %0 = memref.get_global @__constant_1xi8 : memref<1xi8, 2> // 2 is expected memory space
   return
 }

Unfortunately, bufferization throws error,

error: 'memref.global' op initial value expected to be of type 'tensor<1xi8>', but was of type 'tensor<1xi8, 2 : i64>'

it looks like a bug of MLIR. ← need help to confirm.
if it is not a bug. ← need help to solve my problem.

4

I think this fixes the problem: [mlir][memref]: Fix Bug in GlobalOp Verifier by FranklandJack · Pull Request #144900 · llvm/llvm-project · GitHub

1 Like