[MLIR] Inconsistent output when executing MLIR program with and without `-test-loop-fusion=test-loop-fusion-transformation`

[Lambor24]

My git version is ea44647.

Description:

I am experiencing an inconsistent result when executing the same MLIR program with and without the -test-loop-fusion=test-loop-fusion-transformation.

Steps to Reproduce:

1. MLIR Program (test.mlir):

test.mlir:

module {
  func.func private @printMemrefI64(tensor<*xi64>)
  func.func @main() -> () {
    %arg0 = "tosa.const"() {value = dense<20> : tensor<36x2x3xi64>} : () -> tensor<36x2x3xi64>
    %0 = tosa.reduce_sum %arg0 {axis = 0 : i32} : (tensor<36x2x3xi64>) -> tensor<1x2x3xi64>
    %rtn1 = tensor.cast %0 : tensor<1x2x3xi64> to tensor<*xi64>
    call @printMemrefI64(%rtn1) : (tensor<*xi64>) -> ()
    return
  }
}

2. Command to Run Without -test-loop-fusion=test-loop-fusion-transformation:

/path/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline='builtin.module(func.func(tosa-to-linalg-named,tosa-to-linalg))' | \
/path/llvm-project/build/bin/mlir-opt -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-affine-loops -lower-affine -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -finalize-memref-to-llvm -convert-func-to-llvm -reconcile-unrealized-casts | \
/path/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void \
-shared-libs=/path/llvm-project/build/lib/libmlir_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_c_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_async_runtime.so

3. Output Without -test-loop-fusion=test-loop-fusion-transformation:

[[[720,    720,    720], 
  [720,    720,    720]]]

4. Command to Run With -test-loop-fusion=test-loop-fusion-transformation:

/path/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline='builtin.module(func.func(tosa-to-linalg-named,tosa-to-linalg))' | \
/path/llvm-project/build/bin/mlir-opt -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-affine-loops -test-loop-fusion=test-loop-fusion-transformation -lower-affine -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -finalize-memref-to-llvm -convert-func-to-llvm -reconcile-unrealized-casts | \
/path/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void \
-shared-libs=/path/llvm-project/build/lib/libmlir_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_c_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_async_runtime.so

5. Output With -test-loop-fusion=test-loop-fusion-transformation:

[[[20,    20,    20], 
  [20,    20,    20]]]

I'm not sure if there is any bug in my program or if the wrong usage of the above passes caused this result.

[ftynse]

This is a test pass intended for testing. It doesn't have any promise of correctness.

[joker-eph]

I see it as a design issue with the test pass: I believe test passes should promise correctness on un-annotated IR: that is, if there is no legality check in the utility exercised by the pass, then the pass should use some user-input to assert where to apply the transformation, here for example the pass could look for loops annotated with an attribute indicating what to fuse, instead of just assuming that any loop in the IR should be fused.

[bondhugula]

I see it as a design issue with the test pass: I believe test passes should promise correctness on un-annotated IR: that is, if there is no legality check in the utility exercised by the pass, then the pass should use some user-input to assert where to apply the transformation, here for example the pass could look for loops annotated with an attribute indicating what to fuse, instead of just assuming that any loop in the IR should be fused.

This is not a test pass meant to test fusion, but to test some of the utilities used by loop fusion and only those utilities. The test pass does not and will not have all the checks and other code needed to perform fusion, nor is it intended to test a larger number of other utilities used by the main pass (those utilities and code is tested by the main pass's test cases themselves). So the expectation of valid fused output here is incorrect. I've renamed the pass option to test-loop-fusion-utilities to avoid confusion. The output being generated is the correct output that the fuseLoops utility is expected to generate.
#148320

[llvmbot]

@llvm/issue-subscribers-mlir-affine

Author: None (Lambor24)

My git version is [ea44647](https://github.com/llvm/llvm-project/commit/ea44647a0b49de826191eeb6e05020262b5a81e9).

Description:

I am experiencing an inconsistent result when executing the same MLIR program with and without the -test-loop-fusion=test-loop-fusion-transformation.

Steps to Reproduce:

1. MLIR Program (test.mlir):

test.mlir:

module {
  func.func private @<!-- -->printMemrefI64(tensor&lt;*xi64&gt;)
  func.func @<!-- -->main() -&gt; () {
    %arg0 = "tosa.const"() {value = dense&lt;20&gt; : tensor&lt;36x2x3xi64&gt;} : () -&gt; tensor&lt;36x2x3xi64&gt;
    %0 = tosa.reduce_sum %arg0 {axis = 0 : i32} : (tensor&lt;36x2x3xi64&gt;) -&gt; tensor&lt;1x2x3xi64&gt;
    %rtn1 = tensor.cast %0 : tensor&lt;1x2x3xi64&gt; to tensor&lt;*xi64&gt;
    call @<!-- -->printMemrefI64(%rtn1) : (tensor&lt;*xi64&gt;) -&gt; ()
    return
  }
}

2. Command to Run Without -test-loop-fusion=test-loop-fusion-transformation:

/path/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline='builtin.module(func.func(tosa-to-linalg-named,tosa-to-linalg))' | \
/path/llvm-project/build/bin/mlir-opt -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-affine-loops -lower-affine -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -finalize-memref-to-llvm -convert-func-to-llvm -reconcile-unrealized-casts | \
/path/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void \
-shared-libs=/path/llvm-project/build/lib/libmlir_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_c_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_async_runtime.so

3. Output Without -test-loop-fusion=test-loop-fusion-transformation:

[[[720,    720,    720], 
  [720,    720,    720]]]

4. Command to Run With -test-loop-fusion=test-loop-fusion-transformation:

/path/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline='builtin.module(func.func(tosa-to-linalg-named,tosa-to-linalg))' | \
/path/llvm-project/build/bin/mlir-opt -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-affine-loops -test-loop-fusion=test-loop-fusion-transformation -lower-affine -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -finalize-memref-to-llvm -convert-func-to-llvm -reconcile-unrealized-casts | \
/path/llvm-project/build/bin/mlir-cpu-runner -e main -entry-point-result=void \
-shared-libs=/path/llvm-project/build/lib/libmlir_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_c_runner_utils.so \
-shared-libs=/path/llvm-project/build/lib/libmlir_async_runtime.so

5. Output With -test-loop-fusion=test-loop-fusion-transformation:

[[[20,    20,    20], 
  [20,    20,    20]]]

I'm not sure if there is any bug in my program or if the wrong usage of the above passes caused this result.