[MLIR] Inconsistent output when executing MLIR program with and without -affine-loop-fusion

[Lambor24]

My git version is 6003c30.

Description:

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

Steps to Reproduce:

1. MLIR Program (test.mlir):

test.mlir:

module {
  func.func private @printMemrefF32(tensor<*xf32>)
  func.func @main() {
    %0 = "tosa.const"() <{values = dense<2.000000e+00> : tensor<1x3x9x4xf32>}> : () -> tensor<1x3x9x4xf32>
    %1 = "tosa.const"() <{values = dense<3.000000e+00> : tensor<12x2x5x4xf32>}> : () -> tensor<12x2x5x4xf32>
    %2 = "tosa.const"() <{values = dense<4.000000e+00> : tensor<12xf32>}> : () -> tensor<12xf32>
    %3 = "tosa.const"() <{values = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32>
    %4 = "tosa.const"() <{values = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32>
    %5 = tosa.conv2d %0, %1, %2, %3, %4 {acc_type = f32, dilation = array<i64: 1, 2>, pad = array<i64: 1, 1, 1, 1>, stride = array<i64: 1, 1>} : (tensor<1x3x9x4xf32>, tensor<12x2x5x4xf32>, tensor<12xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x4x3x12xf32>
    %6 = tosa.reduce_sum %5 {axis = 1 : i32} : (tensor<1x4x3x12xf32>) -> tensor<1x1x3x12xf32>
    %cast = tensor.cast %6 : tensor<1x1x3x12xf32> to tensor<*xf32>
    call @printMemrefF32(%cast) : (tensor<*xf32>) -> ()
    return
  }
}

2. Command to Run Without -affine-loop-fusion:

/path/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline='builtin.module(func.func(tosa-to-linalg-named))' | \
/path/llvm-project/build/bin/mlir-opt -pass-pipeline='builtin.module(func.func(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 -convert-math-to-llvm -canonicalize -finalize-memref-to-llvm -convert-func-to-llvm | \
/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 -affine-loop-fusion:

[[[[592,     592,     592,     592,     592,     592,     592,     592,     592,     592,     592,     592], 
   [736,     736,     736,     736,     736,     736,     736,     736,     736,     736,     736,     736], 
   [592,     592,     592,     592,     592,     592,     592,     592,     592,     592,     592,     592]]]]

4. Command to Run With -affine-loop-fusion:

/path/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline='builtin.module(func.func(tosa-to-linalg-named))' | \
/path/llvm-project/build/bin/mlir-opt -pass-pipeline='builtin.module(func.func(tosa-to-linalg))' | \
/path/llvm-project/build/bin/mlir-opt -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-affine-loops -affine-loop-fusion -lower-affine -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -convert-math-to-llvm -canonicalize -finalize-memref-to-llvm -convert-func-to-llvm | \
/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 -affine-loop-fusion:

[[[[16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16], 
   [16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16], 
   [16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16]]]]

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

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 -affine-loop-fusion.

Steps to Reproduce:

1. MLIR Program (test.mlir):

test.mlir:

module {
  func.func private @<!-- -->printMemrefF32(tensor&lt;*xf32&gt;)
  func.func @<!-- -->main() {
    %0 = "tosa.const"() &lt;{values = dense&lt;2.000000e+00&gt; : tensor&lt;1x3x9x4xf32&gt;}&gt; : () -&gt; tensor&lt;1x3x9x4xf32&gt;
    %1 = "tosa.const"() &lt;{values = dense&lt;3.000000e+00&gt; : tensor&lt;12x2x5x4xf32&gt;}&gt; : () -&gt; tensor&lt;12x2x5x4xf32&gt;
    %2 = "tosa.const"() &lt;{values = dense&lt;4.000000e+00&gt; : tensor&lt;12xf32&gt;}&gt; : () -&gt; tensor&lt;12xf32&gt;
    %3 = "tosa.const"() &lt;{values = dense&lt;0.000000e+00&gt; : tensor&lt;1xf32&gt;}&gt; : () -&gt; tensor&lt;1xf32&gt;
    %4 = "tosa.const"() &lt;{values = dense&lt;0.000000e+00&gt; : tensor&lt;1xf32&gt;}&gt; : () -&gt; tensor&lt;1xf32&gt;
    %5 = tosa.conv2d %0, %1, %2, %3, %4 {acc_type = f32, dilation = array&lt;i64: 1, 2&gt;, pad = array&lt;i64: 1, 1, 1, 1&gt;, stride = array&lt;i64: 1, 1&gt;} : (tensor&lt;1x3x9x4xf32&gt;, tensor&lt;12x2x5x4xf32&gt;, tensor&lt;12xf32&gt;, tensor&lt;1xf32&gt;, tensor&lt;1xf32&gt;) -&gt; tensor&lt;1x4x3x12xf32&gt;
    %6 = tosa.reduce_sum %5 {axis = 1 : i32} : (tensor&lt;1x4x3x12xf32&gt;) -&gt; tensor&lt;1x1x3x12xf32&gt;
    %cast = tensor.cast %6 : tensor&lt;1x1x3x12xf32&gt; to tensor&lt;*xf32&gt;
    call @<!-- -->printMemrefF32(%cast) : (tensor&lt;*xf32&gt;) -&gt; ()
    return
  }
}

2. Command to Run Without -affine-loop-fusion:

/path/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline='builtin.module(func.func(tosa-to-linalg-named))' | \
/path/llvm-project/build/bin/mlir-opt -pass-pipeline='builtin.module(func.func(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 -convert-math-to-llvm -canonicalize -finalize-memref-to-llvm -convert-func-to-llvm | \
/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 -affine-loop-fusion:

[[[[592,     592,     592,     592,     592,     592,     592,     592,     592,     592,     592,     592], 
   [736,     736,     736,     736,     736,     736,     736,     736,     736,     736,     736,     736], 
   [592,     592,     592,     592,     592,     592,     592,     592,     592,     592,     592,     592]]]]

4. Command to Run With -affine-loop-fusion:

/path/llvm-project/build/bin/mlir-opt test.mlir -pass-pipeline='builtin.module(func.func(tosa-to-linalg-named))' | \
/path/llvm-project/build/bin/mlir-opt -pass-pipeline='builtin.module(func.func(tosa-to-linalg))' | \
/path/llvm-project/build/bin/mlir-opt -tosa-to-arith -one-shot-bufferize="bufferize-function-boundaries" -convert-linalg-to-affine-loops -affine-loop-fusion -lower-affine -convert-scf-to-cf -expand-strided-metadata -convert-cf-to-llvm -convert-arith-to-llvm -convert-math-to-llvm -canonicalize -finalize-memref-to-llvm -convert-func-to-llvm | \
/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 -affine-loop-fusion:

[[[[16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16], 
   [16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16], 
   [16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16,     16]]]]

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