[MLIR] Have some conceptual issues with the MLIR framework

MLIR
1

I am a beginner familiarizing myself with the MLIR project, and recently I’ve encountered some issues understanding its structure which can be listed below：

  1. currently the kinds of regions contained in mlir is SSACFG and Graph , I am quite familiar with the architecture of SSA (Static Single Assignment) and can understand its structure, after I read the documentation GraphRegion, the doc says that this region is appropriate for concurrent semantics without control flow , but it doesn’t provide specific examples ，so I’m a little confused about this concept
  2. the doc gives an example:
"test.graph_region"() ({ // A Graph region
  %1 = "op1"(%1, %3) : (i32, i32) -> (i32)  // OK: %1, %3 allowed here
  %2 = "test.ssacfg_region"() ({
     %5 = "op2"(%1, %2, %3, %4) : (i32, i32, i32, i32) -> (i32) // OK: %1, %2, %3, %4 all defined in the containing region
  }) : () -> (i32)
  %3 = "op2"(%1, %4) : (i32, i32) -> (i32)  // OK: %4 allowed here
  %4 = "op3"(%1) : (i32) -> (i32)
}) : () -> ()

The SSA region is nested within the graph region， How do these two regions combine logically? I don’t quite understand.

I would really appreciate any help you can provide. :slight_smile:

2

There is no CFG control flow in the graph region in the example, but if/when the control is transferred to the nested region, these you find again the usual control flow.

It’s hard to find a concrete example that you could relate to without knowing your background, but imagine that the outer region is a ML dataflow graph, and there you could have a “custom operation” which defines a nested region implemented in LLVM dialect directly.

3

Thank a lot ! After reading the answer and revisiting the documentation, I have a clearer understanding of the concept of graph regions.
And another question bothers me: I think structure like LLVM using Module->Function->Block is OK, so why mlir use: operation->region->block , and the block also contains operation ,region, so what is the purpose of this design

4

The previous example should apply here as well?

Imagine a module, that contains a graph, this graph defines a graph region to model a ML dataflow graph, and there you could have a “custom operation” which defines a nested region implemented in LLVM dialect directly.

Now you have module>graph>custom_op>…

Another example can be the way we can model host code and device code using GPU module.

module {
  gpu.module {
    gpu.func @kernel { ...}
  }
  func @host() { ...}
}
1 Like
5

Got it ! Thanks a lot :blush: