[Mlir-commits] [mlir] [mlir][linalg] Pack matmul pass (PR #89782)

[Andrzej Warzyński]

================

----------------
banach-space wrote:

Given that there's no other documentation attached to this PR, I'm using these tests to infer the high level logic. That's a bit tricky ATM - it could be simplified with more meaningful var names (MLIR and LIT) ;-) 

For example:

```mlir
func.func @block_matmul(%lhs: tensor<128x128xf32>, %rhs: tensor<128x128xf32>, %out: tensor<128x128xf32>) -> tensor<128x128xf32> {
  %0 = linalg.matmul  ins(%lhs, %rhs : tensor<128x128xf32>, tensor<128x128xf32>)
                      outs(%our : tensor<128x128xf32>) -> tensor<128x128xf32>
  return %0 : tensor<128x128xf32>
}
```
or
```
func.func @block_matmul(%A: tensor<128x128xf32>, %B: tensor<128x128xf32>, %C: tensor<128x128xf32>) -> tensor<128x128xf32> {
  %0 = linalg.matmul  ins(%A, %B : tensor<128x128xf32>, tensor<128x128xf32>)
                      outs(%C : tensor<128x128xf32>) -> tensor<128x128xf32>
  return %0 : tensor<128x128xf32>
}
```

And then in LIT:
```mlir
// CHECK: %[[LHS_PACKED:.+]] = tensor.pack %[[LHS]]
```

Or something similar. I don't really mind what the actual names are, as long as they help to map to the high level semantics of Matmul. Updated names will really help to review this, thanks!

https://github.com/llvm/llvm-project/pull/89782