[WIP] Dump option to print CuTE layouts for scheduled Fusion

[jacobhinkle]

No description provided.

[github-actions]

Description

• Add CuTE layout printing for scheduled fusion

• Introduce printCute() to dump Fusion IR in CuTE format

• Implement CuteConverter for translating IterDomains to CuTE layouts

• Support symbolic strides and shapes via MultipliedString and IntTuple

---

Changes walkthrough 📝

	Relevant files
Enhancement	7 files lower2device.cpp Enable CuTE layout dumping in debug mode                                  +3/-0      fusion.cpp Implement printCute for CuTE layout output                              +66/-0    cute_translation.cpp Implement CuTE layout conversion logic                                      +305/-0  tensor_view.cpp Add fullName method for TensorView                                              +9/-3      fusion.h Declare printCute method for Fusion                                            +10/-0    cute_translation.h Define CuTE layout translation structures                                +240/-0  interface_nodes.h Declare fullName in TensorView                                                      +2/-0
Configuration changes	3 files options.cpp Add FusionIrCute debug dump option                                              +1/-0      options.h Add FusionIrCute to DebugDumpOption                                            +1/-0      CMakeLists.txt Include cute_translation.cpp in build                                        +1/-0

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review Unimplemented Support The Swizzle2D operation is not supported and throws an exception, which may lead to incomplete or incorrect CuTE layout generation when such operations are present in the fusion. } else if (auto* swizzle = dynamic_cast<Swizzle2D*>(expr)) { // TODO NVF_THROW("Swizzle2D support not yet implemented ", swizzle->toString()); Incomplete Merge Handling The handling of merge operations in the getLayout function does not check for contiguity and always represents merges as IntTuples, potentially missing optimization opportunities for contiguous merges. // TODO: Check if this is a contiguous merge. If so, then we can flatten // it IF NOT contiguous merge, then create a new IntTuple consisting of // the incoming shapes/strides // For now, we don't flatten any merges, we just represent them as // IntTuple always Int out_shape = std::make_shared<IntTuple>( std::vector<Int>{outer_shape, inner_shape}); Int out_stride = std::make_shared<IntTuple>( std::vector<Int>{outer_stride, inner_stride}); id_size_stride.emplace( merge->out(), std::pair<Int, Int>{out_shape, out_stride}); } else if (auto* swizzle = dynamic_cast<Swizzle*>(expr)) { Unhandled Multiplication Case The multiplication of two IntTuples in the operator* function for Int is not properly handled and defaults to a string representation, which may result in incorrect stride calculations. // TODO: Handle multiplication of IntTuples properly return {MultipliedString{a.toString() + "*" + b.toString()}};