[mlir][SPIR-V] Question: How to lower math.isFinite to SPIR-V?

[Rayfxl]

Hi, I'm working on a project that involves lowering math.isFinite to SPIR-V. I've noticed a few things and would like to confirm whether the current infrastructure supports this, and how best to proceed.

Current observations:

1. math.isFinite is already supported and lowered to the LLVM and NVVM dialects.

2. However, I couldn't find any existing support for lowering math.isFinite to the SPIR-V dialect.

3. The SPIR-V dialect itself does not define an equivalent operation (e.g., spirv.IsFinite), but the SPIR-V specification does include the OpIsFinite instruction.

My questions:

1. Is it expected that math.isFinite should eventually lower to a SPIR-V operation like OpIsFinite?

2. If so, should a new spirv.IsFinite op be introduced in the SPIR-V dialect to support this?

3. Or is there a recommended workaround for this case (e.g. combine isinf/isnan with boolean logic)?

I'm currently exploring the feasibility of this lowering and not yet planning to submit a patch. I’d appreciate any guidance on the intended direction for this feature.

Thanks!

[llvmbot]

@llvm/issue-subscribers-mlir

Author: None (Rayfxl)

Hi, I'm working on a project that involves lowering math.isFinite to SPIR-V. I've noticed a few things and would like to confirm whether the current infrastructure supports this, and how best to proceed.

Current observations:

1. math.isFinite is already supported and lowered to the LLVM and NVVM dialects.

2. However, I couldn't find any existing support for lowering math.isFinite to the SPIR-V dialect.

3. The SPIR-V dialect itself does not define an equivalent operation (e.g., spirv.IsFinite), but the SPIR-V specification does include the OpIsFinite instruction.

My questions:

1. Is it expected that math.isFinite should eventually lower to a SPIR-V operation like OpIsFinite?

2. If so, should a new spirv.IsFinite op be introduced in the SPIR-V dialect to support this?

3. Or is there a recommended workaround for this case (e.g. combine isinf/isnan with boolean logic)?

I'm currently exploring the feasibility of this lowering and not yet planning to submit a patch. I’d appreciate any guidance on the intended direction for this feature.

Thanks!

[llvmbot]

@llvm/issue-subscribers-mlir-spirv

Author: Xiaolei Feng (Rayfxl)

Hi, I'm working on a project that involves lowering math.isFinite to SPIR-V. I've noticed a few things and would like to confirm whether the current infrastructure supports this, and how best to proceed.

Current observations:

1. math.isFinite is already supported and lowered to the LLVM and NVVM dialects.

2. However, I couldn't find any existing support for lowering math.isFinite to the SPIR-V dialect.

3. The SPIR-V dialect itself does not define an equivalent operation (e.g., spirv.IsFinite), but the SPIR-V specification does include the OpIsFinite instruction.

My questions:

1. Is it expected that math.isFinite should eventually lower to a SPIR-V operation like OpIsFinite?

2. If so, should a new spirv.IsFinite op be introduced in the SPIR-V dialect to support this?

3. Or is there a recommended workaround for this case (e.g. combine isinf/isnan with boolean logic)?

I'm currently exploring the feasibility of this lowering and not yet planning to submit a patch. I’d appreciate any guidance on the intended direction for this feature.

Thanks!

[IgWod-IMG]

Hey,

1. Is it expected that math.isFinite should eventually lower to a SPIR-V operation like OpIsFinite?

In my opinion, yes. The simplest lowering is probably the best here.

2. If so, should a new spirv.IsFinite op be introduced in the SPIR-V dialect to support this?

Again in my opinion: yes, we should introduce a new op based on OpIsFinite.

3. Or is there a recommended workaround for this case (e.g. combine isinf/isnan with boolean logic)?

I'd say any workaround is less desirable if we have an op in SPIR-V we can directly lower into.

[Rayfxl]

Hey,

1. Is it expected that math.isFinite should eventually lower to a SPIR-V operation like OpIsFinite?

In my opinion, yes. The simplest lowering is probably the best here.

2. If so, should a new spirv.IsFinite op be introduced in the SPIR-V dialect to support this?

Again in my opinion: yes, we should introduce a new op based on OpIsFinite.

3. Or is there a recommended workaround for this case (e.g. combine isinf/isnan with boolean logic)?

I'd say any workaround is less desirable if we have an op in SPIR-V we can directly lower into.

Thanks for the clarification!

I'd like to move forward with adding spirv.IsFinite to the SPIR-V dialect and lowering math.isFinite to it. I saw that the dialect provides the define_inst.sh script in utils/spirv, so I’d like to ask a few questions before proceeding.

1. Using the script

• After running define_inst.sh, what parts of the generated op typically need to be updated manually? (e.g., traits, argument/result types, etc.) I looked at similar ops like IsNan, and it seems there might be little to update.
• The documentation also mentions writing a custom parser/printer or verifier. Would that be expected or necessary for OpIsFinite?

2. Manual alternative

• If the script can’t be used (e.g., due to limited scripting or offline environments), is it acceptable to manually define the op in .td files?
• What would be the equivalent manual workflow to follow if not using the script?

3. Lowering

• For lowering math.isFinite, would using CheckedElementwiseOpPattern<math::IsFiniteOp, spirv::IsFiniteOp> be acceptable if the op interfaces match?
• Or would a hand-written OpConversionPattern be preferred?

4. Reference ops

• Are there existing SPIR-V ops you'd recommend as references for both op definition and lowering?
• Ideally something unary and float-based.

Thanks for your guidance!

[kuhar]

You can add it by hand, our automatic import scripts are mainly meant for bulk import of large extensions etc. I'd base the implementation off IsNan/IsInf

[Rayfxl]

You can add it by hand, our automatic import scripts are mainly meant for bulk import of large extensions etc. I'd base the implementation off IsNan/IsInf

Thanks for the helpful responses!

I'll be implementing the spirv.IsFinite op and the lowering from math.isFinite,
and plan to submit a PR later on.

Will link it here once it's ready.