[MLIR] Incorrect fusion pass behavior with `-test-loop-fusion=test-loop-fusion-transformation`

[Lambor24]

My git version is 6003c30.

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 {
  memref.global "private" constant @__constant_5x1x2x3xi16 : memref<5x1x2x3xi16> = dense<81> {alignment = 64 : i64}
  func.func private @printMemrefI16(memref<*xi16>) attributes {llvm.emit_c_interface}
  func.func @main() {
    %c81_i16 = arith.constant 81 : i16
    %c1_i16 = arith.constant 1 : i16
    %alloc = memref.alloc() {alignment = 64 : i64} : memref<1x2x3xi16>
    affine.for %arg0 = 0 to 1 {
      affine.for %arg1 = 0 to 2 {
        affine.for %arg2 = 0 to 3 {
          affine.store %c1_i16, %alloc[%arg0, %arg1, %arg2] : memref<1x2x3xi16>
        }
      }
    }
    affine.for %arg0 = 0 to 5 {
      affine.for %arg1 = 0 to 1 {
        affine.for %arg2 = 0 to 2 {
          affine.for %arg3 = 0 to 3 {
            %0 = affine.load %alloc[%arg1, %arg2, %arg3] : memref<1x2x3xi16>
            %1 = arith.muli %0, %c81_i16 : i16
            affine.store %1, %alloc[%arg1, %arg2, %arg3] : memref<1x2x3xi16>
          }
        }
      }
    }
    %expand_shape = memref.expand_shape %alloc [[0, 1], [2], [3]] output_shape [1, 1, 2, 3] : memref<1x2x3xi16> into memref<1x1x2x3xi16>
    %cast = memref.cast %expand_shape : memref<1x1x2x3xi16> to memref<*xi16>
    call @printMemrefI16(%cast) : (memref<*xi16>) -> ()
    return
  }
}

output:

module {
  memref.global "private" constant @__constant_5x1x2x3xi16 : memref<5x1x2x3xi16> = dense<81> {alignment = 64 : i64}
  func.func private @printMemrefI16(memref<*xi16>) attributes {llvm.emit_c_interface}
  func.func @main() {
    %c0 = arith.constant 0 : index
    %c81_i16 = arith.constant 81 : i16
    %c1_i16 = arith.constant 1 : i16
    %alloc = memref.alloc() {alignment = 64 : i64} : memref<1x2x3xi16>
    affine.for %arg0 = 0 to 5 {
      affine.for %arg1 = 0 to 1 {
        affine.for %arg2 = 0 to 2 {
          affine.for %arg3 = 0 to 3 {
            affine.store %c1_i16, %alloc[%c0, %arg2, %arg3] : memref<1x2x3xi16>
            %0 = affine.load %alloc[%arg1, %arg2, %arg3] : memref<1x2x3xi16>
            %1 = arith.muli %0, %c81_i16 : i16
            affine.store %1, %alloc[%arg1, %arg2, %arg3] : memref<1x2x3xi16>
          }
        }
      }
    }
    %expand_shape = memref.expand_shape %alloc [[0, 1], [2], [3]] output_shape [1, 1, 2, 3] : memref<1x2x3xi16> into memref<1x1x2x3xi16>
    %cast = memref.cast %expand_shape : memref<1x1x2x3xi16> to memref<*xi16>
    call @printMemrefI16(%cast) : (memref<*xi16>) -> ()
    return
  }
}

Obviously, the two nested loops here should not fuse together.

[llvmbot]

@llvm/issue-subscribers-mlir

Author: None (Lambor24)

My git version is [6003c30](https://github.com/llvm/llvm-project/commit/6003c3055a4630be31cc3d459cdbb88248a007b9).

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 {
  memref.global "private" constant @<!-- -->__constant_5x1x2x3xi16 : memref&lt;5x1x2x3xi16&gt; = dense&lt;81&gt; {alignment = 64 : i64}
  func.func private @<!-- -->printMemrefI16(memref&lt;*xi16&gt;) attributes {llvm.emit_c_interface}
  func.func @<!-- -->main() {
    %c81_i16 = arith.constant 81 : i16
    %c1_i16 = arith.constant 1 : i16
    %alloc = memref.alloc() {alignment = 64 : i64} : memref&lt;1x2x3xi16&gt;
    affine.for %arg0 = 0 to 1 {
      affine.for %arg1 = 0 to 2 {
        affine.for %arg2 = 0 to 3 {
          affine.store %c1_i16, %alloc[%arg0, %arg1, %arg2] : memref&lt;1x2x3xi16&gt;
        }
      }
    }
    affine.for %arg0 = 0 to 5 {
      affine.for %arg1 = 0 to 1 {
        affine.for %arg2 = 0 to 2 {
          affine.for %arg3 = 0 to 3 {
            %0 = affine.load %alloc[%arg1, %arg2, %arg3] : memref&lt;1x2x3xi16&gt;
            %1 = arith.muli %0, %c81_i16 : i16
            affine.store %1, %alloc[%arg1, %arg2, %arg3] : memref&lt;1x2x3xi16&gt;
          }
        }
      }
    }
    %expand_shape = memref.expand_shape %alloc [[0, 1], [2], [3]] output_shape [1, 1, 2, 3] : memref&lt;1x2x3xi16&gt; into memref&lt;1x1x2x3xi16&gt;
    %cast = memref.cast %expand_shape : memref&lt;1x1x2x3xi16&gt; to memref&lt;*xi16&gt;
    call @<!-- -->printMemrefI16(%cast) : (memref&lt;*xi16&gt;) -&gt; ()
    return
  }
}

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

/path/llvm-project/build/bin/mlir-opt test.mlir -lower-affine -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -convert-math-to-llvm -finalize-memref-to-llvm -convert-func-to-llvm -reconcile-unrealized-casts | \
/path/llvm-project/build/bin/mlir-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:

[[[[7057,     7057,     7057], 
   [7057,     7057,     7057]]]]

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

/path/llvm-project/build/bin/mlir-opt test.mlir -test-loop-fusion=test-loop-fusion-transformation -lower-affine -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -convert-math-to-llvm -finalize-memref-to-llvm -convert-func-to-llvm -reconcile-unrealized-casts | \
/path/llvm-project/build/bin/mlir-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:

[[[[81,     81,     81], 
   [81,     81,     81]]]]

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

[llvmbot]

@llvm/issue-subscribers-mlir-affine

Author: None (Lambor24)

My git version is [6003c30](https://github.com/llvm/llvm-project/commit/6003c3055a4630be31cc3d459cdbb88248a007b9).

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 {
  memref.global "private" constant @<!-- -->__constant_5x1x2x3xi16 : memref&lt;5x1x2x3xi16&gt; = dense&lt;81&gt; {alignment = 64 : i64}
  func.func private @<!-- -->printMemrefI16(memref&lt;*xi16&gt;) attributes {llvm.emit_c_interface}
  func.func @<!-- -->main() {
    %c81_i16 = arith.constant 81 : i16
    %c1_i16 = arith.constant 1 : i16
    %alloc = memref.alloc() {alignment = 64 : i64} : memref&lt;1x2x3xi16&gt;
    affine.for %arg0 = 0 to 1 {
      affine.for %arg1 = 0 to 2 {
        affine.for %arg2 = 0 to 3 {
          affine.store %c1_i16, %alloc[%arg0, %arg1, %arg2] : memref&lt;1x2x3xi16&gt;
        }
      }
    }
    affine.for %arg0 = 0 to 5 {
      affine.for %arg1 = 0 to 1 {
        affine.for %arg2 = 0 to 2 {
          affine.for %arg3 = 0 to 3 {
            %0 = affine.load %alloc[%arg1, %arg2, %arg3] : memref&lt;1x2x3xi16&gt;
            %1 = arith.muli %0, %c81_i16 : i16
            affine.store %1, %alloc[%arg1, %arg2, %arg3] : memref&lt;1x2x3xi16&gt;
          }
        }
      }
    }
    %expand_shape = memref.expand_shape %alloc [[0, 1], [2], [3]] output_shape [1, 1, 2, 3] : memref&lt;1x2x3xi16&gt; into memref&lt;1x1x2x3xi16&gt;
    %cast = memref.cast %expand_shape : memref&lt;1x1x2x3xi16&gt; to memref&lt;*xi16&gt;
    call @<!-- -->printMemrefI16(%cast) : (memref&lt;*xi16&gt;) -&gt; ()
    return
  }
}

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

/path/llvm-project/build/bin/mlir-opt test.mlir -lower-affine -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -convert-math-to-llvm -finalize-memref-to-llvm -convert-func-to-llvm -reconcile-unrealized-casts | \
/path/llvm-project/build/bin/mlir-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:

[[[[7057,     7057,     7057], 
   [7057,     7057,     7057]]]]

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

/path/llvm-project/build/bin/mlir-opt test.mlir -test-loop-fusion=test-loop-fusion-transformation -lower-affine -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -convert-math-to-llvm -finalize-memref-to-llvm -convert-func-to-llvm -reconcile-unrealized-casts | \
/path/llvm-project/build/bin/mlir-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:

[[[[81,     81,     81], 
   [81,     81,     81]]]]

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

[bondhugula]

A test pass isn't meant to be used in a pipeline for semantically valid transformations and to generate correct code. If there is a bug in affine fusion itself, please post the input and output for just the fusion pass.

[joker-eph]

I believe test passes should be written to be "safe". That does not mean they need complex analyses, they can get whatever information they need to drive them from an attribute for example and just do nothing on random IR.

[lipracer]

After reviewing the relevant tests for this pass, it seems that the original unit tests were also incorrect.It seems that the count of the loop has not been checked either.

[bondhugula]

I believe test passes should be written to be "safe". That does not mean they need complex analyses, they can get whatever information they need to drive them from an attribute for example and just do nothing on random IR.

Sure.

[bondhugula]

@lipracer are you planning to fix this issue? If not, we should reassign. Also, please paste the output before and after the test fusion pass instead of just an integration test, which doesn't establish or say why the output of the test-fusion pass isn't the expected one.

[lipracer]

This issue is actually due to incomplete functionality. Please refer to mlir/include/mlir/Dialect/Affine/LoopFusionUtils. h:110 for details. I was not very clear about the original design, so I gave up fixing it.Actually, there are similar issues in the testing:

llvm-project/mlir/test/Dialect/Affine/loop-fusion-transformation.mlir

Line 7 in 035dcf6

affine.for %i0 = 0 to 16 {