[MLIR] Unexpected inf output from `tosa.rsqrt` on zero input

[sweead]

test commit: 9164d20

Description:

When applying tosa.rsqrt to a tensor containing the value 0.0, the resulting value becomes inf.

Steps to Reproduce:

Minimal MLIR program (test.mlir):

module {
  func.func private @printMemrefF32(tensor<*xf32>)
  func.func @main() -> () {
    %arg0 = "tosa.const"() {values = dense<0.0> : tensor<1xf32>} : () -> tensor<1xf32>
    %3 = tosa.rsqrt %arg0 : (tensor<1xf32>) -> tensor<1xf32>
    %rtn1 = tensor.cast %3 : tensor<1xf32> to tensor<*xf32>
    call @printMemrefF32(%rtn1) : (tensor<*xf32>) -> ()
    return
  }
}

Command:

mlir-opt test.mlir  --tosa-to-linalg-pipeline  --sparsifier |  \
mlir-runner -e main -entry-point-result=void -shared-libs=/home/workdir/llvm-project/build/lib/libmlir_runner_utils.so

Output:

[inf]

[llvmbot]

@llvm/issue-subscribers-mlir

Author: Anonymous_reporter (sweead)

test commit: [9164d20](https://github.com/llvm/llvm-project/commit/9164d206b33d61c93f5fc4628797485f96d654ca)

Description:

When applying tosa.rsqrt to a tensor containing the value 0.0, the resulting value becomes inf.

Steps to Reproduce:

Minimal MLIR program (test.mlir):

module {
  func.func private @<!-- -->printMemrefF32(tensor&lt;*xf32&gt;)
  func.func @<!-- -->main() -&gt; () {
    %arg0 = "tosa.const"() {values = dense&lt;0.0&gt; : tensor&lt;1xf32&gt;} : () -&gt; tensor&lt;1xf32&gt;
    %3 = tosa.rsqrt %arg0 : (tensor&lt;1xf32&gt;) -&gt; tensor&lt;1xf32&gt;
    %rtn1 = tensor.cast %3 : tensor&lt;1xf32&gt; to tensor&lt;*xf32&gt;
    call @<!-- -->printMemrefF32(%rtn1) : (tensor&lt;*xf32&gt;) -&gt; ()
    return
  }
}

Command:

mlir-opt test.mlir  --tosa-to-linalg-pipeline  --sparsifier |  \
mlir-runner -e main -entry-point-result=void -shared-libs=/home/workdir/llvm-project/build/lib/libmlir_runner_utils.so

Output:

[inf]

[psunn]

the behavior of tosa.rsqrt returning inf for an input of 0.0 is mathematically correct in floating-point arithmetic, also matching the Floating-point behavior of rsqrt from TOSA Spec.

[sweead]

the behavior of tosa.rsqrt returning inf for an input of 0.0 is mathematically correct in floating-point arithmetic, also matching the Floating-point behavior of rsqrt from TOSA Spec.

Thank you! I understand now