[MetaXGPU] Use ties-to-even rounding for tirx.round on the MACA target

[Lfan-ke]

Problem

MACAMath lowers tirx.round to round / roundf, which round halfway cases away from zero. Every other backend in the tree lowers it to nearbyint, which under the default rounding mode rounds halfway cases to even:

• src/backend/cuda/codegen/intrin_rule_cuda.cc:41-45 returns nearbyint for fp64 and nearbyintf for fp32
• src/target/llvm/intrin_rule_llvm.cc:91-93 dispatches tirx.round to llvm::Intrinsic::nearbyint

So the same graph rounds 0.5 to 1.0 on a MACA device and to 0.0 everywhere else, and a value that gets constant-folded at compile time disagrees with the same value computed on device. Only exact halfway inputs diverge, so this stays invisible under random test data.

fp16 is already correct: it maps to hrint, which is ties-to-even.

Changes

MACAMath returns nearbyint for fp64 and nearbyintf for fp32, matching the CUDA rule it was forked from. tirx.nearbyint is unaffected — it already lowered to the same function.

Test Plan

tests/python/codegen/test_target_codegen_maca_round.py on a MetaX C500 (MACA 3.5.3.20). Both cases feed the eight halfway values ±0.5, ±1.5, ±2.5, ±3.5, which are the only inputs the two rules disagree on:

• the generated MACA source calls nearbyintf and does not call roundf
• the kernel returns 0, 2, 2, 4, -0, -2, -2, -4

Test Result

2 passed in 1.19s

Reverting the two lines in intrin_rule_maca.cc and rebuilding turns both cases red, so the test does exercise the changed lowering:

Mismatched elements: 4 / 8 (50%)
 [0]: 1.0 (ACTUAL), 0.0 (DESIRED)
 [2]: 3.0 (ACTUAL), 2.0 (DESIRED)
 [4]: -1.0 (ACTUAL), -0.0 (DESIRED)
 [6]: -3.0 (ACTUAL), -2.0 (DESIRED)

[gemini-code-assist]

Code Review

This pull request updates the MACA math codegen rules in intrin_rule_maca.cc to map the 'round' function to 'nearbyint' (for 64-bit floats) and 'nearbyintf' (for 32-bit floats) to use ties-to-even rounding, matching constant-folding semantics. There are no review comments, and I have no feedback to provide.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[Lfan-ke]

@codex review please. read code at pr and repositories.