commit 8da4564c4689f529d05611887423077eaacbfd01
author Nirvedh Meshram <96096277+nirvedhmeshram@users.noreply.github.com> Thu Aug 22 18:49:29 2024 -0500
committer GitHub <noreply@github.com> Thu Aug 22 23:49:29 2024 +0000
tree d357de0664b60482e2e4e61fcbc4180f7ef71a6f
parent 588732ce4402618aa6b339228efbbee2d35f222a

Revert "[Flow] Convert from tensor.cast to flow.tensor.reshape early …(#18256)" (#18331)

This reverts commit 1c0c5a6ff64bedb1cb1275eaba16aab3fc26acdf.

This is causing an issue with
https://github.com/iree-org/iree/actions/runs/10505447157/job/29119827242#step:5:2269
iree/tests/e2e/regression/check_regression_llvm-cpu_layernorm.mlir

Triaging the bug I see that one of the dispatches is slightly different
but should result to the same numerics but it does not
Here is how the problem dispatch used to be originally
```
        func.func @old_dispatch2(%11 : tensor<128x384xf32>, %12: tensor<128xf32>) -> tensor<128x384xf32> {
        %cst = arith.constant 5.000000e+00 : f32
        %cst_0 = arith.constant 0.000000e+00 : f32
        %13 = tensor.empty() : tensor<128x384xf32>
        %14 = tensor.empty() : tensor<128xf32>
        %15 = linalg.fill ins(%cst_0 : f32) outs(%14 : tensor<128xf32>) -> tensor<128xf32>
        %16 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0)>], iterator_types = ["parallel", "reduction"]} ins(%11 : tensor<128x384xf32>) outs(%15 : tensor<128xf32>) {
        ^bb0(%in: f32, %out: f32):
          %18 = arith.addf %in, %out : f32
          linalg.yield %18 : f32
        } -> tensor<128xf32>
        %17 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0)>, affine_map<(d0, d1) -> (d0)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%11, %16, %12 : tensor<128x384xf32>, tensor<128xf32>, tensor<128xf32>) outs(%13 : tensor<128x384xf32>) {
        ^bb0(%og_in : f32, %in: f32, %in_1: f32, %out: f32):
          %18 = arith.mulf %in, %in_1 : f32
          %19 = arith.subf %og_in, %18 : f32
          linalg.yield %19 : f32
        } -> tensor<128x384xf32>
        return %17 :  tensor<128x384xf32>
        }
  ```
  with the reverting PR this dispatch is becoming
  ```
          func.func @new_dispatch2(%11 : tensor<128x384xf32>, %12: tensor<128xf32>) -> tensor<128x384xf32> {
        %cst = arith.constant 5.000000e+00 : f32
        %cst_0 = arith.constant 0.000000e+00 : f32
        %13 = tensor.empty() : tensor<128x384xf32>
        %14 = tensor.empty() : tensor<128xf32>
        %15 = linalg.fill ins(%cst_0 : f32) outs(%14 : tensor<128xf32>) -> tensor<128xf32>
        %16 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0)>], iterator_types = ["parallel", "reduction"]} ins(%11 : tensor<128x384xf32>) outs(%15 : tensor<128xf32>) {
        ^bb0(%in: f32, %out: f32):
          %18 = arith.addf %in, %out : f32
          linalg.yield %18 : f32
        } -> tensor<128xf32>
        %17 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0)>, affine_map<(d0, d1) -> (d0)>, affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} ins(%16, %12 : tensor<128xf32>, tensor<128xf32>) outs(%13 : tensor<128x384xf32>) {
        ^bb0(%in: f32, %in_1: f32, %out: f32):
          %18 = arith.mulf %in, %in_1 : f32
          %19 = arith.subf %cst, %18 : f32
          linalg.yield %19 : f32
        } -> tensor<128x384xf32>
        return %17 :  tensor<128x384xf32>
        }
   ```
   The difference is at `%19 = arith.subf %cst, %18 : f32`
   Note that in both cases `%11 : tensor<128x384xf32>` is `5.0` from the model input and hence the output *should* be same, however on `arch64` it is not but on x86 it is, the IR at the mlir llvm dialect is identicalbetween x86 and arch64 so its not some easy to  spot intrinsics / fast math kind of bug AFAICS