[mlir] Inconsistent output when executing MLIR program with `--scf-parallel-loop-tiling` `--canonicalize`

[Emilyaxe]

git version: b2b267e

system: Ubuntu 18.04.6 LTS

Description:

I am experiencing an inconsistent result when executing the same MLIR program with and without --scf-parallel-loop-tiling --canonicalize.
The output becomes correct when either of these two options is removed, so I'm unsure which optimization contains the bug.

Steps to Reproduce:

1. MLIR Program (a.mlir):

a.mlir:

module {
  func.func private @printMemrefI32(tensor<*xi32>)
  func.func private @printMemrefF32(tensor<*xf32>)
  func.func @main() {
    %7 = "tosa.const"() <{values = dense<6220> : tensor<1x6x6xi32>}> : () -> tensor<1x6x6xi32>
    %9 = "tosa.const"() <{values = dense<-298> : tensor<1x6x6xi32>}> : () -> tensor<1x6x6xi32>
    %51 = tosa.bitwise_or %7, %9 : (tensor<1x6x6xi32>, tensor<1x6x6xi32>) -> tensor<1x6x6xi32>
    %cast = tensor.cast %51 : tensor<1x6x6xi32> to tensor<*xi32>
    call @printMemrefI32(%cast) : (tensor<*xi32>) -> ()
    return
  }
}

2. Command to Run without optimizations :

/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt a.mlir -pass-pipeline="builtin.module(func.func(tosa-to-linalg))" | /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt  \
 -tosa-to-arith  -one-shot-bufferize="bufferize-function-boundaries"    -convert-linalg-to-parallel-loops   \
 -convert-index-to-llvm      -convert-arith-to-llvm  -convert-scf-to-cf       -convert-arith-to-llvm    \
  -convert-cf-to-llvm  -finalize-memref-to-llvm   -convert-func-to-llvm  -lower-affine  -convert-arith-to-llvm \
  -reconcile-unrealized-casts | timeout 10 /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-cpu-runner -e main \
-entry-point-result=void --shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_c_runner_utils.so \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_runner_utils.so  \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_async_runtime.so

3. Output without optimizations ::

[[[-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290]]]

4. Command to Run with optimizations :

/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt a.mlir -pass-pipeline="builtin.module(func.func(tosa-to-linalg))" | /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt \
-tosa-to-arith  -one-shot-bufferize="bufferize-function-boundaries" \
-convert-linalg-to-parallel-loops            -convert-index-to-llvm      -convert-arith-to-llvm    \
--scf-parallel-loop-tiling="parallel-loop-tile-sizes=1,4 no-min-max-bounds=true"    \
--canonicalize  --scf-parallel-loop-tiling="parallel-loop-tile-sizes=1,4 no-min-max-bounds=true"  \
 -convert-scf-to-cf       -convert-arith-to-llvm       -convert-cf-to-llvm  -finalize-memref-to-llvm  \
 -convert-func-to-llvm  -lower-affine  -convert-arith-to-llvm  \
-reconcile-unrealized-casts | timeout 10 /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void  \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_c_runner_utils.so \
 --shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_runner_utils.so  \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_async_runtime.so

5. Output with optimizations :

[[[-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [0,    0,    0,    32767,    -1553697280,    22067], 
  [-1553697280,    22067,    -1553697264,    22067,    0,    0]]]

[Emilyaxe]

I have analyzed the issue and found that the root cause may be related to the --scf-parallel-loop-tiling="parallel-loop-tile-sizes=1,4 no-min-max-bounds=true" option.
The input IR (after running --canonicalize) input.txt can correctly print the index %10 (from 0 to 5). However the output IR (after running --scf-parallel-loop-tiling="parallel-loop-tile-sizes=1,4 no-min-max-bounds=true") output.txt
cannot correctly print the index %46 (from 0 to 3), which leads to random values in the last two rows of the final results. I am not pretty sure whether the use of no-min-max-bounds=true is appropriate in this case.

Running the input.txt

with commands

/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt-953838d input.txt \
-convert-scf-to-cf       -convert-arith-to-llvm       -convert-cf-to-llvm -convert-vector-to-llvm -finalize-memref-to-llvm  \
-convert-func-to-llvm  -lower-affine  -convert-arith-to-llvm  \
-reconcile-unrealized-casts | timeout 10 /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void  \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_c_runner_utils.so \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_runner_utils.so  

we will get results as following:

0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
4
5
4
5
4
5
4
5
4
5
4
5
Unranked Memref base@ = 0x56475a1f8100 rank = 3 offset = 0 sizes = [1, 6, 6] strides = [36, 6, 1] data = 
[[[-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290]]]

with commands

/data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-opt-953838d input.txt \
-convert-scf-to-cf       -convert-arith-to-llvm       -convert-cf-to-llvm -convert-vector-to-llvm -finalize-memref-to-llvm  \
-convert-func-to-llvm  -lower-affine  -convert-arith-to-llvm  \
-reconcile-unrealized-casts | timeout 10 /data/szy/MLIR/llvm-release/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void  \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_c_runner_utils.so \
--shared-libs=/data/szy/MLIR/llvm-release/llvm-project/build/lib/libmlir_runner_utils.so  

However, running the output.txt with the same commands as the one used in running input.txt

we will get results as following:

0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
0
1
2
3
Unranked Memref base@ = 0x5631d802b440 rank = 3 offset = 0 sizes = [1, 6, 6] strides = [36, 6, 1] data = 
[[[-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [-290,    -290,    -290,    -290,    -290,    -290], 
  [2,    0,    0,    0,    0,    0], 
  [0,    0,    0,    1879426263,    852879496,    0]]]

This case is a little complex, and I haven't been able to pinpoint the exact invalid IR in output.txt yet. If further analysis is needed, I will continue investigating.
Hi, @linearhit @sherhut @bondhugula
Sorry to disturb you, but could you please kindly take a moment to review this problem? Thank you!